| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from contextlib import AsyncExitStack, asynccontextmanager |
| 5 | from types import SimpleNamespace |
| 6 | from unittest.mock import AsyncMock, MagicMock |
| 7 | |
| 8 | import pytest |
| 9 | import pytest_asyncio |
| 10 | from openai.types.responses import ( |
| 11 | ResponseFunctionToolCall, |
| 12 | ResponseOutputItemDoneEvent, |
| 13 | ResponseOutputMessage, |
| 14 | ResponseOutputText, |
| 15 | ResponseReasoningItem, |
| 16 | ResponseReasoningTextDeltaEvent, |
| 17 | ResponseReasoningTextDoneEvent, |
| 18 | ResponseTextConfig, |
| 19 | ResponseTextDeltaEvent, |
| 20 | ) |
| 21 | from openai.types.responses import ( |
| 22 | ResponseUsage as OpenAIResponseUsage, |
| 23 | ) |
| 24 | from openai.types.responses.response_format_text_json_schema_config import ( |
| 25 | ResponseFormatTextJSONSchemaConfig, |
| 26 | ) |
| 27 | from openai.types.responses.tool import ( |
| 28 | CodeInterpreterContainerCodeInterpreterToolAuto, |
| 29 | LocalShell, |
| 30 | Mcp, |
| 31 | Tool, |
| 32 | ) |
| 33 | from openai_harmony import Message as OpenAIHarmonyMessage |
| 34 | from openai_harmony import Role |
| 35 | |
| 36 | import vllm.envs as envs |
| 37 | from tests.entrypoints.openai.chat_completion.test_serving_chat import MockModelConfig |
| 38 | from vllm.entrypoints.generate.base.protocol import ( |
| 39 | DeltaFunctionCall, |
| 40 | DeltaMessage, |
| 41 | DeltaToolCall, |
| 42 | RequestResponseMetadata, |
| 43 | ) |
| 44 | from vllm.entrypoints.mcp.tool_server import ToolServer |
| 45 | from vllm.entrypoints.openai.responses.context import ( |
| 46 | ConversationContext, |
| 47 | HarmonyContext, |
| 48 | ParsableContext, |
| 49 | SimpleContext, |
| 50 | ) |
| 51 | from vllm.entrypoints.openai.responses.protocol import ( |
| 52 | ResponseCompletedEvent, |
| 53 | ResponseCreatedEvent, |
| 54 | ResponseRawMessageAndToken, |
| 55 | ResponsesRequest, |
| 56 | ResponsesResponse, |
| 57 | serialize_message, |
| 58 | ) |
| 59 | from vllm.entrypoints.openai.responses.serving import ( |
| 60 | OpenAIServingResponses, |
| 61 | extract_tool_types, |
| 62 | ) |
| 63 | from vllm.entrypoints.openai.responses.streaming_events import ( |
| 64 | StreamingState, |
| 65 | ) |
| 66 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 67 | from vllm.inputs import tokens_input |
| 68 | from vllm.logprobs import Logprob as SampleLogprob |
| 69 | from vllm.outputs import CompletionOutput, RequestOutput |
| 70 | from vllm.parser.harmony import Segment |
| 71 | from vllm.renderers import TokenizeParams |
| 72 | from vllm.renderers.online_renderer import ( |
| 73 | OnlineRenderer, |
| 74 | _extract_allowed_tools_from_mcp_requests, |
| 75 | ) |
| 76 | from vllm.sampling_params import SamplingParams |
| 77 | from vllm.v1.metrics.stats import RequestStateStats |
| 78 | |
| 79 | pytestmark = pytest.mark.skip_global_cleanup |
| 80 | |
| 81 | |
| 82 | def _new_online_renderer() -> OnlineRenderer: |
| 83 | renderer = OnlineRenderer.__new__(OnlineRenderer) |
| 84 | renderer.exclude_tools_when_tool_choice_none = False |
| 85 | renderer.parser = None |
| 86 | renderer.trust_request_chat_template = True |
| 87 | return renderer |
| 88 | |
| 89 | |
| 90 | class MockConversationContext(ConversationContext): |
| 91 | """Mock conversation context for testing.""" |
| 92 | |
| 93 | def __init__(self): |
| 94 | self.init_tool_sessions_called = False |
| 95 | self.init_tool_sessions_args = None |
| 96 | self.init_tool_sessions_kwargs = None |
| 97 | |
| 98 | def append_output(self, output) -> None: |
| 99 | pass |
| 100 | |
| 101 | def append_tool_output(self, output) -> None: |
| 102 | pass |
| 103 | |
| 104 | async def call_tool(self): |
| 105 | return [] |
| 106 | |
| 107 | def need_builtin_tool_call(self) -> bool: |
| 108 | return False |
| 109 | |
| 110 | def render_for_completion(self): |
| 111 | return [] |
| 112 | |
| 113 | async def init_tool_sessions(self, tool_server, exit_stack, request_id, mcp_tools): |
| 114 | self.init_tool_sessions_called = True |
| 115 | self.init_tool_sessions_args = (tool_server, exit_stack, request_id, mcp_tools) |
| 116 | |
| 117 | async def cleanup_session(self) -> None: |
| 118 | pass |
| 119 | |
| 120 | |
| 121 | def test_serialize_message_pydantic_model_returns_dict() -> None: |
| 122 | msg = ResponseRawMessageAndToken(message="hello", tokens=[1, 2, 3]) |
| 123 | |
| 124 | serialized = serialize_message(msg) |
| 125 | |
| 126 | assert isinstance(serialized, dict) |
| 127 | assert serialized["type"] == "raw_message_tokens" |
| 128 | assert serialized["message"] == "hello" |
| 129 | |
| 130 | |
| 131 | @pytest.fixture |
| 132 | def mock_context(): |
| 133 | """Create a mock conversation context.""" |
| 134 | return MockConversationContext() |
| 135 | |
| 136 | |
| 137 | @pytest.fixture |
| 138 | def mock_exit_stack(): |
| 139 | """Create a mock async exit stack.""" |
| 140 | return MagicMock(spec=AsyncExitStack) |
| 141 | |
| 142 | |
| 143 | def test_extract_tool_types(monkeypatch: pytest.MonkeyPatch) -> None: |
| 144 | tools: list[Tool] = [] |
| 145 | assert extract_tool_types(tools) == set() |
| 146 | |
| 147 | tools.append(LocalShell(type="local_shell")) |
| 148 | assert extract_tool_types(tools) == {"local_shell"} |
| 149 | |
| 150 | tools.append(CodeInterpreterContainerCodeInterpreterToolAuto(type="auto")) |
| 151 | assert extract_tool_types(tools) == {"local_shell", "auto"} |
| 152 | |
| 153 | tools.extend( |
| 154 | [ |
| 155 | Mcp(type="mcp", server_label="random", server_url=""), |
| 156 | Mcp(type="mcp", server_label="container", server_url=""), |
| 157 | Mcp(type="mcp", server_label="code_interpreter", server_url=""), |
| 158 | Mcp(type="mcp", server_label="web_search_preview", server_url=""), |
| 159 | ] |
| 160 | ) |
| 161 | # When envs.VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS is not set, |
| 162 | # mcp tool types are all ignored. |
| 163 | assert extract_tool_types(tools) == {"local_shell", "auto"} |
| 164 | |
| 165 | # container is allowed, it would be extracted |
| 166 | monkeypatch.setenv("VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS", "container") |
| 167 | assert extract_tool_types(tools) == {"local_shell", "auto", "container"} |
| 168 | |
| 169 | # code_interpreter and web_search_preview are allowed, |
| 170 | # they would be extracted |
| 171 | monkeypatch.setenv( |
| 172 | "VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS", "code_interpreter,web_search_preview" |
| 173 | ) |
| 174 | assert extract_tool_types(tools) == { |
| 175 | "local_shell", |
| 176 | "auto", |
| 177 | "code_interpreter", |
| 178 | "web_search_preview", |
| 179 | } |
| 180 | |
| 181 | |
| 182 | @pytest.mark.skip_global_cleanup |
| 183 | def test_response_created_event_uses_public_json_schema_alias() -> None: |
| 184 | schema = { |
| 185 | "type": "object", |
| 186 | "properties": { |
| 187 | "event_name": {"type": "string"}, |
| 188 | "date": {"type": "string"}, |
| 189 | "participants": {"type": "array", "items": {"type": "string"}}, |
| 190 | }, |
| 191 | "required": ["event_name", "date", "participants"], |
| 192 | "additionalProperties": False, |
| 193 | } |
| 194 | text = ResponseTextConfig() |
| 195 | text.format = ResponseFormatTextJSONSchemaConfig( |
| 196 | type="json_schema", |
| 197 | name="calendar_event", |
| 198 | schema=schema, |
| 199 | description="A calendar event.", |
| 200 | strict=True, |
| 201 | ) |
| 202 | request = ResponsesRequest( |
| 203 | model="test-model", |
| 204 | input="Alice and Bob are going to a science fair on Friday.", |
| 205 | text=text, |
| 206 | ) |
| 207 | sampling_params = request.to_sampling_params(default_max_tokens=64) |
| 208 | initial_response = ResponsesResponse.from_request( |
| 209 | request=request, |
| 210 | sampling_params=sampling_params, |
| 211 | model_name="test-model", |
| 212 | created_time=0, |
| 213 | output=[], |
| 214 | status="in_progress", |
| 215 | usage=None, |
| 216 | ).model_dump(mode="json", by_alias=True) |
| 217 | |
| 218 | fmt = initial_response["text"]["format"] |
| 219 | assert fmt["schema"] == schema |
| 220 | assert "schema_" not in fmt |
| 221 | |
| 222 | event = ResponseCreatedEvent( |
| 223 | type="response.created", |
| 224 | sequence_number=0, |
| 225 | response=initial_response, |
| 226 | ) |
| 227 | assert event.response.text is not None |
| 228 | assert event.response.text.format is not None |
| 229 | assert event.response.text.format.model_dump(by_alias=True)["schema"] == schema |
| 230 | |
| 231 | |
| 232 | @pytest.mark.asyncio |
| 233 | async def test_online_renderer_renders_non_harmony_responses_with_explicit_history(): |
| 234 | renderer = _new_online_renderer() |
| 235 | renderer.use_harmony = False |
| 236 | renderer.chat_template = "server-template" |
| 237 | renderer.chat_template_content_format = "string" |
| 238 | renderer.default_chat_template_kwargs = {"server_default": "kept"} |
| 239 | renderer.parser = None |
| 240 | renderer.preprocess_chat = AsyncMock( |
| 241 | return_value=( |
| 242 | [], |
| 243 | [tokens_input([11, 12, 13], cache_salt="request-salt")], |
| 244 | ) |
| 245 | ) |
| 246 | previous_messages = [ |
| 247 | {"role": "system", "content": "old instructions"}, |
| 248 | {"role": "user", "content": "first"}, |
| 249 | ] |
| 250 | previous_outputs = [ |
| 251 | ResponseOutputMessage( |
| 252 | id="msg_1", |
| 253 | content=[ |
| 254 | ResponseOutputText( |
| 255 | annotations=[], |
| 256 | text="prior answer", |
| 257 | type="output_text", |
| 258 | ) |
| 259 | ], |
| 260 | role="assistant", |
| 261 | status="completed", |
| 262 | type="message", |
| 263 | ) |
| 264 | ] |
| 265 | request = ResponsesRequest( |
| 266 | input="next", |
| 267 | instructions="new instructions", |
| 268 | tools=[], |
| 269 | cache_salt="request-salt", |
| 270 | chat_template_kwargs={"request_default": "kept"}, |
| 271 | ) |
| 272 | |
| 273 | result = await renderer.render_responses( |
| 274 | request, |
| 275 | previous_messages=previous_messages, |
| 276 | previous_response_outputs=previous_outputs, |
| 277 | ) |
| 278 | |
| 279 | assert not isinstance(result, ErrorResponse) |
| 280 | assert result.messages == [ |
| 281 | {"role": "system", "content": "new instructions"}, |
| 282 | {"role": "user", "content": "first"}, |
| 283 | {"role": "assistant", "content": "prior answer"}, |
| 284 | {"role": "user", "content": "next"}, |
| 285 | ] |
| 286 | assert result.engine_input["prompt_token_ids"] == [11, 12, 13] |
| 287 | assert result.engine_input["cache_salt"] == "request-salt" |
| 288 | assert previous_messages[0]["content"] == "old instructions" |
| 289 | assert previous_outputs[0].content[0].text == "prior answer" |
| 290 | |
| 291 | preprocess_kwargs = renderer.preprocess_chat.call_args.kwargs |
| 292 | assert preprocess_kwargs["default_template"] == "server-template" |
| 293 | assert preprocess_kwargs["default_template_content_format"] == "string" |
| 294 | assert preprocess_kwargs["default_template_kwargs"]["server_default"] == "kept" |
| 295 | assert preprocess_kwargs["default_template_kwargs"]["request_default"] == "kept" |
| 296 | |
| 297 | |
| 298 | _REUSED_IDS = [10, 20, 30] |
| 299 | _IMAGE_PART = { |
| 300 | "type": "input_image", |
| 301 | "image_url": "https://example.com/a.png", |
| 302 | "detail": "auto", |
| 303 | } |
| 304 | |
| 305 | |
| 306 | async def _render_responses_with_reuse(request, previous_messages=None): |
| 307 | online_renderer = OnlineRenderer( |
| 308 | model_config=MockModelConfig(), |
| 309 | renderer=MagicMock(), |
| 310 | request_logger=None, |
| 311 | chat_template=None, |
| 312 | chat_template_content_format="auto", |
| 313 | ) |
| 314 | render_chat = AsyncMock(return_value=([[]], [tokens_input([1, 2])])) |
| 315 | online_renderer.renderer.render_chat_async = render_chat |
| 316 | |
| 317 | result = await online_renderer.render_responses( |
| 318 | request, previous_messages=previous_messages |
| 319 | ) |
| 320 | |
| 321 | assert not isinstance(result, ErrorResponse) |
| 322 | return render_chat, result.engine_input |
| 323 | |
| 324 | |
| 325 | @pytest.mark.asyncio |
| 326 | @pytest.mark.parametrize("ids", [_REUSED_IDS, [1.5]]) |
| 327 | @pytest.mark.parametrize( |
| 328 | "request_input", |
| 329 | [ |
| 330 | [ |
| 331 | { |
| 332 | "role": "user", |
| 333 | "content": [{"type": "input_text", "text": "look"}, _IMAGE_PART], |
| 334 | } |
| 335 | ], |
| 336 | [ |
| 337 | { |
| 338 | "type": "function_call", |
| 339 | "call_id": "call_1", |
| 340 | "name": "shot", |
| 341 | "arguments": "{}", |
| 342 | }, |
| 343 | { |
| 344 | "type": "function_call_output", |
| 345 | "call_id": "call_1", |
| 346 | "output": [_IMAGE_PART], |
| 347 | }, |
| 348 | ], |
| 349 | ], |
| 350 | ids=["input_image", "tool_output_image"], |
| 351 | ) |
| 352 | async def test_responses_kv_transfer_prompt_token_ids_ignored_with_non_text_input( |
| 353 | request_input, ids |
| 354 | ): |
| 355 | request = ResponsesRequest( |
| 356 | input=request_input, |
| 357 | kv_transfer_params={"do_remote_prefill": True, "prompt_token_ids": ids}, |
| 358 | ) |
| 359 | |
| 360 | render_chat, _ = await _render_responses_with_reuse(request) |
| 361 | |
| 362 | render_chat.assert_awaited_once() |
| 363 | assert request.kv_transfer_params == {"do_remote_prefill": True} |
| 364 | |
| 365 | |
| 366 | @pytest.mark.asyncio |
| 367 | async def test_responses_kv_transfer_prompt_token_ids_ignored_with_media_in_history(): |
| 368 | request = ResponsesRequest( |
| 369 | input="next", kv_transfer_params={"prompt_token_ids": _REUSED_IDS} |
| 370 | ) |
| 371 | |
| 372 | render_chat, _ = await _render_responses_with_reuse( |
| 373 | request, previous_messages=[{"role": "user", "content": [_IMAGE_PART]}] |
| 374 | ) |
| 375 | |
| 376 | render_chat.assert_awaited_once() |
| 377 | |
| 378 | |
| 379 | @pytest.mark.asyncio |
| 380 | async def test_responses_kv_transfer_prompt_token_ids_kept_with_text_input(): |
| 381 | request = ResponsesRequest( |
| 382 | input=[ |
| 383 | {"role": "user", "content": [{"type": "input_text", "text": "hi"}]}, |
| 384 | {"role": "assistant", "content": "ok"}, |
| 385 | ], |
| 386 | kv_transfer_params={"prompt_token_ids": _REUSED_IDS}, |
| 387 | ) |
| 388 | |
| 389 | _, engine_input = await _render_responses_with_reuse(request) |
| 390 | |
| 391 | assert engine_input["prompt_token_ids"] == _REUSED_IDS |
| 392 | |
| 393 | |
| 394 | @pytest.mark.asyncio |
| 395 | async def test_online_renderer_rejects_untrusted_responses_chat_template(): |
| 396 | renderer = _new_online_renderer() |
| 397 | renderer.use_harmony = False |
| 398 | renderer.chat_template = "server-template" |
| 399 | renderer.chat_template_content_format = "string" |
| 400 | renderer.default_chat_template_kwargs = {} |
| 401 | renderer.exclude_tools_when_tool_choice_none = False |
| 402 | renderer.trust_request_chat_template = False |
| 403 | renderer.parser = None |
| 404 | renderer.preprocess_chat = AsyncMock( |
| 405 | return_value=([], [tokens_input([1])]), |
| 406 | ) |
| 407 | request = ResponsesRequest( |
| 408 | input="hello", |
| 409 | chat_template_kwargs={"chat_template": "{{ messages }}"}, |
| 410 | ) |
| 411 | |
| 412 | result = await renderer.render_responses(request) |
| 413 | |
| 414 | assert isinstance(result, ErrorResponse) |
| 415 | assert result.error.code == 400 |
| 416 | assert "untrusted chat template" in result.error.message.lower() |
| 417 | renderer.preprocess_chat.assert_not_awaited() |
| 418 | |
| 419 | |
| 420 | @pytest.mark.asyncio |
| 421 | async def test_online_renderer_rejects_harmony_history_for_non_harmony_model(): |
| 422 | renderer = _new_online_renderer() |
| 423 | renderer.use_harmony = False |
| 424 | request = ResponsesRequest(input="next") |
| 425 | |
| 426 | result = await renderer.render_responses( |
| 427 | request, |
| 428 | previous_messages=[ |
| 429 | OpenAIHarmonyMessage.from_role_and_content(Role.USER, "first") |
| 430 | ], |
| 431 | ) |
| 432 | |
| 433 | assert isinstance(result, ErrorResponse) |
| 434 | assert result.error.param == "previous_response_id" |
| 435 | |
| 436 | |
| 437 | @pytest.mark.asyncio |
| 438 | async def test_online_renderer_rejects_chat_history_for_harmony_model(): |
| 439 | renderer = _new_online_renderer() |
| 440 | renderer.use_harmony = True |
| 441 | request = ResponsesRequest(input="next") |
| 442 | |
| 443 | result = await renderer.render_responses( |
| 444 | request, |
| 445 | previous_messages=[{"role": "user", "content": "first"}], |
| 446 | ) |
| 447 | |
| 448 | assert isinstance(result, ErrorResponse) |
| 449 | assert result.error.param == "previous_response_id" |
| 450 | |
| 451 | |
| 452 | @pytest.mark.asyncio |
| 453 | async def test_online_renderer_adjusts_harmony_tool_choice( |
| 454 | monkeypatch: pytest.MonkeyPatch, |
| 455 | ): |
| 456 | class StubParser: |
| 457 | def __init__(self, tokenizer, tools, *, model_config): |
| 458 | pass |
| 459 | |
| 460 | def adjust_request(self, request): |
| 461 | request.cache_salt = "adjusted" |
| 462 | return request |
| 463 | |
| 464 | renderer = _new_online_renderer() |
| 465 | renderer.use_harmony = True |
| 466 | renderer.parser = StubParser |
| 467 | renderer.model_config = SimpleNamespace(max_model_len=100) |
| 468 | tokenizer = SimpleNamespace(truncation_side="left") |
| 469 | renderer.renderer = SimpleNamespace( |
| 470 | tokenizer=tokenizer, |
| 471 | get_tokenizer=lambda: tokenizer, |
| 472 | ) |
| 473 | monkeypatch.setattr( |
| 474 | "vllm.renderers.online_renderer.render_for_completion", |
| 475 | lambda messages: [1], |
| 476 | ) |
| 477 | request = ResponsesRequest( |
| 478 | input="next", |
| 479 | tools=[ |
| 480 | { |
| 481 | "type": "function", |
| 482 | "name": "lookup", |
| 483 | "parameters": {"type": "object", "properties": {}}, |
| 484 | } |
| 485 | ], |
| 486 | tool_choice={"type": "function", "name": "lookup"}, |
| 487 | ) |
| 488 | |
| 489 | result = await renderer.render_responses(request) |
| 490 | |
| 491 | assert not isinstance(result, ErrorResponse) |
| 492 | assert result.engine_input["cache_salt"] == "adjusted" |
| 493 | |
| 494 | |
| 495 | @pytest.mark.asyncio |
| 496 | async def test_online_renderer_applies_responses_token_budget_to_harmony_prompt( |
| 497 | monkeypatch: pytest.MonkeyPatch, |
| 498 | ): |
| 499 | renderer = _new_online_renderer() |
| 500 | renderer.use_harmony = True |
| 501 | renderer.model_config = SimpleNamespace(max_model_len=5) |
| 502 | renderer.renderer = SimpleNamespace( |
| 503 | tokenizer=SimpleNamespace(truncation_side="left") |
| 504 | ) |
| 505 | monkeypatch.setattr( |
| 506 | "vllm.renderers.online_renderer.render_for_completion", |
| 507 | lambda messages: [1, 2, 3, 4, 5, 6], |
| 508 | ) |
| 509 | request = ResponsesRequest( |
| 510 | input="hello", |
| 511 | max_output_tokens=2, |
| 512 | truncation="auto", |
| 513 | ) |
| 514 | |
| 515 | result = await renderer.render_responses(request) |
| 516 | |
| 517 | assert not isinstance(result, ErrorResponse) |
| 518 | assert result.engine_input["prompt_token_ids"] == [4, 5, 6] |
| 519 | |
| 520 | |
| 521 | @pytest.mark.asyncio |
| 522 | async def test_online_renderer_renders_harmony_continuation_with_call_linkage( |
| 523 | monkeypatch: pytest.MonkeyPatch, |
| 524 | ): |
| 525 | renderer = _new_online_renderer() |
| 526 | renderer.use_harmony = True |
| 527 | renderer.model_config = SimpleNamespace(max_model_len=100) |
| 528 | renderer.renderer = SimpleNamespace( |
| 529 | tokenizer=SimpleNamespace(truncation_side="left") |
| 530 | ) |
| 531 | monkeypatch.setattr( |
| 532 | "vllm.renderers.online_renderer.render_for_completion", |
| 533 | lambda messages: [1], |
| 534 | ) |
| 535 | previous_messages = [ |
| 536 | OpenAIHarmonyMessage.from_role_and_content(Role.USER, "first"), |
| 537 | ] |
| 538 | previous_outputs = [ |
| 539 | ResponseFunctionToolCall( |
| 540 | arguments='{"city":"Paris"}', |
| 541 | call_id="call_1", |
| 542 | name="lookup_weather", |
| 543 | type="function_call", |
| 544 | ) |
| 545 | ] |
| 546 | request = ResponsesRequest( |
| 547 | input=[ |
| 548 | { |
| 549 | "type": "function_call_output", |
| 550 | "call_id": "call_1", |
| 551 | "output": "sunny", |
| 552 | } |
| 553 | ], |
| 554 | instructions="replacement instructions are ignored on continuations", |
| 555 | reasoning={"effort": "high"}, |
| 556 | tools=[ |
| 557 | { |
| 558 | "type": "function", |
| 559 | "name": "lookup_weather", |
| 560 | "parameters": {"type": "object", "properties": {}}, |
| 561 | } |
| 562 | ], |
| 563 | cache_salt="request-salt", |
| 564 | ) |
| 565 | |
| 566 | result = await renderer.render_responses( |
| 567 | request, |
| 568 | previous_messages=previous_messages, |
| 569 | previous_response_outputs=previous_outputs, |
| 570 | ) |
| 571 | |
| 572 | assert not isinstance(result, ErrorResponse) |
| 573 | assert result.messages[0] is previous_messages[0] |
| 574 | tool_output = result.messages[-1] |
| 575 | assert tool_output.author.role == "tool" |
| 576 | assert tool_output.author.name == "functions.lookup_weather" |
| 577 | assert tool_output.channel == "commentary" |
| 578 | assert tool_output.recipient == "assistant" |
| 579 | assert result.engine_input["cache_salt"] == "request-salt" |
| 580 | assert "arrival_time" in result.engine_input |
| 581 | assert previous_messages == [ |
| 582 | OpenAIHarmonyMessage.from_role_and_content(Role.USER, "first"), |
| 583 | ] |
| 584 | assert previous_outputs[0].call_id == "call_1" |
| 585 | |
| 586 | |
| 587 | @pytest.mark.asyncio |
| 588 | async def test_online_renderer_returns_typed_error_for_bad_harmony_call_reference(): |
| 589 | renderer = _new_online_renderer() |
| 590 | renderer.use_harmony = True |
| 591 | request = ResponsesRequest( |
| 592 | input=[ |
| 593 | { |
| 594 | "type": "function_call_output", |
| 595 | "call_id": "missing", |
| 596 | "output": "orphaned", |
| 597 | } |
| 598 | ], |
| 599 | tools=[], |
| 600 | ) |
| 601 | |
| 602 | result = await renderer.render_responses(request) |
| 603 | |
| 604 | assert isinstance(result, ErrorResponse) |
| 605 | assert result.error.type == "invalid_request_error" |
| 606 | assert result.error.param == "input" |
| 607 | assert "No call message found for missing" in result.error.message |
| 608 | |
| 609 | |
| 610 | @pytest.mark.asyncio |
| 611 | async def test_online_renderer_preserves_missing_harmony_builtin_tool_behavior( |
| 612 | monkeypatch: pytest.MonkeyPatch, |
| 613 | ): |
| 614 | renderer = _new_online_renderer() |
| 615 | renderer.use_harmony = True |
| 616 | renderer.model_config = SimpleNamespace(max_model_len=100) |
| 617 | renderer.renderer = SimpleNamespace( |
| 618 | tokenizer=SimpleNamespace(truncation_side="left") |
| 619 | ) |
| 620 | monkeypatch.setattr( |
| 621 | "vllm.renderers.online_renderer.render_for_completion", |
| 622 | lambda messages: [1], |
| 623 | ) |
| 624 | request = ResponsesRequest( |
| 625 | input="search", |
| 626 | tools=[{"type": "web_search_preview"}], |
| 627 | ) |
| 628 | |
| 629 | result = await renderer.render_responses(request, tool_server=None) |
| 630 | |
| 631 | assert not isinstance(result, ErrorResponse) |
| 632 | |
| 633 | |
| 634 | @pytest.mark.parametrize( |
| 635 | "engine_inputs", |
| 636 | [[], [tokens_input([1]), tokens_input([2])]], |
| 637 | ) |
| 638 | def test_responses_render_result_rejects_non_single_prompt(engine_inputs): |
| 639 | renderer = _new_online_renderer() |
| 640 | |
| 641 | result = renderer._responses_render_result([], engine_inputs) |
| 642 | |
| 643 | assert isinstance(result, ErrorResponse) |
| 644 | assert f"got {len(engine_inputs)}" in result.error.message |
| 645 | |
| 646 | |
| 647 | class TestInitializeToolSessions: |
| 648 | """Test class for _initialize_tool_sessions method.""" |
| 649 | |
| 650 | @pytest.fixture(params=[ParsableContext, HarmonyContext]) |
| 651 | def tool_context(self, request): |
| 652 | context = request.param.__new__(request.param) |
| 653 | context.available_tools = ["browser", "python"] |
| 654 | context._tool_sessions = {} |
| 655 | context.called_tools = set() |
| 656 | return context |
| 657 | |
| 658 | @pytest.mark.asyncio |
| 659 | @pytest.mark.parametrize( |
| 660 | ("available_tools", "called_tools"), |
| 661 | [ |
| 662 | ([], []), |
| 663 | (["browser"], ["browser"]), |
| 664 | (["browser", "python"], []), |
| 665 | (["browser", "python"], ["browser"]), |
| 666 | (["browser", "python"], ["python"]), |
| 667 | (["browser", "python"], ["browser", "python"]), |
| 668 | ], |
| 669 | ids=["empty", "single", "unused", "first", "last", "both"], |
| 670 | ) |
| 671 | async def test_mcp_sessions_cleaned_once_before_close( |
| 672 | self, tool_context, available_tools, called_tools |
| 673 | ): |
| 674 | events = [] |
| 675 | sessions = {} |
| 676 | |
| 677 | @asynccontextmanager |
| 678 | async def new_session(name, *_args): |
| 679 | session = MagicMock() |
| 680 | session.call_tool = AsyncMock( |
| 681 | side_effect=lambda *_: events.append(("cleanup", name)) |
| 682 | ) |
| 683 | sessions[name] = session |
| 684 | try: |
| 685 | yield session |
| 686 | finally: |
| 687 | events.append(("close", name)) |
| 688 | |
| 689 | tool_server = MagicMock(spec=ToolServer) |
| 690 | tool_server.new_session.side_effect = new_session |
| 691 | tool_context.available_tools = available_tools |
| 692 | async with AsyncExitStack() as stack: |
| 693 | await tool_context.init_tool_sessions(tool_server, stack, "req", {}) |
| 694 | # Reinitializing existing sessions must not duplicate cleanup either. |
| 695 | await tool_context.init_tool_sessions(tool_server, stack, "req", {}) |
| 696 | tool_context.called_tools.update(called_tools) |
| 697 | |
| 698 | for name, session in sessions.items(): |
| 699 | if name in called_tools: |
| 700 | session.call_tool.assert_awaited_once_with("cleanup_session", {}) |
| 701 | assert events.index(("cleanup", name)) < events.index(("close", name)) |
| 702 | else: |
| 703 | session.call_tool.assert_not_awaited() |
| 704 | assert [name for event, name in events if event == "close"] == list( |
| 705 | reversed(available_tools) |
| 706 | ) |
| 707 | |
| 708 | @pytest.mark.asyncio |
| 709 | async def test_mcp_partial_initialization_closes_opened_session(self, tool_context): |
| 710 | closed = [] |
| 711 | session = MagicMock(call_tool=AsyncMock()) |
| 712 | |
| 713 | @asynccontextmanager |
| 714 | async def new_session(name, *_args): |
| 715 | if name == "python": |
| 716 | raise RuntimeError("session initialization failed") |
| 717 | try: |
| 718 | yield session |
| 719 | finally: |
| 720 | closed.append(name) |
| 721 | |
| 722 | tool_server = MagicMock(spec=ToolServer) |
| 723 | tool_server.new_session.side_effect = new_session |
| 724 | with pytest.raises(RuntimeError, match="session initialization failed"): |
| 725 | async with AsyncExitStack() as stack: |
| 726 | await tool_context.init_tool_sessions(tool_server, stack, "req", {}) |
| 727 | |
| 728 | assert closed == ["browser"] |
| 729 | session.call_tool.assert_not_awaited() |
| 730 | |
| 731 | @pytest_asyncio.fixture |
| 732 | async def serving_responses_instance(self): |
| 733 | """Create a real OpenAIServingResponses instance for testing.""" |
| 734 | # Create minimal mocks for required dependencies |
| 735 | engine_client = MagicMock() |
| 736 | |
| 737 | model_config = MagicMock() |
| 738 | model_config.max_model_len = 100 |
| 739 | model_config.hf_config.model_type = "test" |
| 740 | model_config.get_diff_sampling_param.return_value = {} |
| 741 | engine_client.model_config = model_config |
| 742 | |
| 743 | engine_client.input_processor = MagicMock() |
| 744 | engine_client.renderer = MagicMock() |
| 745 | |
| 746 | models = MagicMock() |
| 747 | |
| 748 | tool_server = MagicMock(spec=ToolServer) |
| 749 | |
| 750 | # Create the actual instance |
| 751 | instance = OpenAIServingResponses( |
| 752 | engine_client=engine_client, |
| 753 | models=models, |
| 754 | online_renderer=MagicMock(), |
| 755 | request_logger=None, |
| 756 | chat_template=None, |
| 757 | chat_template_content_format="auto", |
| 758 | tool_server=tool_server, |
| 759 | ) |
| 760 | |
| 761 | return instance |
| 762 | |
| 763 | @pytest.mark.asyncio |
| 764 | async def test_stateful_harmony_delegates_named_tool_choice_to_renderer( |
| 765 | self, |
| 766 | serving_responses_instance, |
| 767 | ): |
| 768 | serving_responses_instance.use_harmony = True |
| 769 | serving_responses_instance.online_renderer.render_responses = AsyncMock( |
| 770 | return_value=SimpleNamespace( |
| 771 | messages=[], |
| 772 | engine_input=tokens_input([1]), |
| 773 | ) |
| 774 | ) |
| 775 | request = ResponsesRequest( |
| 776 | input="hello", |
| 777 | tool_choice={"type": "function", "name": "lookup"}, |
| 778 | tools=[ |
| 779 | { |
| 780 | "type": "function", |
| 781 | "name": "lookup", |
| 782 | "parameters": {"type": "object", "properties": {}}, |
| 783 | } |
| 784 | ], |
| 785 | ) |
| 786 | |
| 787 | result = await serving_responses_instance._render_resolved_response_inputs( |
| 788 | request, |
| 789 | None, |
| 790 | ) |
| 791 | |
| 792 | assert result.engine_input["prompt_token_ids"] == [1] |
| 793 | serving_responses_instance.online_renderer.render_responses.assert_awaited_once() |
| 794 | |
| 795 | @pytest.mark.asyncio |
| 796 | async def test_render_next_turn_reuses_shared_responses_renderer( |
| 797 | self, serving_responses_instance |
| 798 | ): |
| 799 | serving_responses_instance.online_renderer.render_responses = AsyncMock( |
| 800 | return_value=SimpleNamespace( |
| 801 | messages=[], |
| 802 | engine_input=tokens_input([8, 9], cache_salt="request-salt"), |
| 803 | ) |
| 804 | ) |
| 805 | serving_responses_instance.online_renderer.preprocess_chat = AsyncMock( |
| 806 | return_value=([], [tokens_input([1])]) |
| 807 | ) |
| 808 | request = ResponsesRequest( |
| 809 | input="first", |
| 810 | instructions="initial instructions", |
| 811 | cache_salt="request-salt", |
| 812 | ) |
| 813 | turn_messages = [ |
| 814 | {"role": "user", "content": "first"}, |
| 815 | {"role": "assistant", "content": "tool call"}, |
| 816 | {"role": "tool", "content": "tool output"}, |
| 817 | ] |
| 818 | |
| 819 | engine_inputs = await serving_responses_instance._render_next_turn( |
| 820 | request, |
| 821 | turn_messages, |
| 822 | ) |
| 823 | |
| 824 | assert engine_inputs[0]["prompt_token_ids"] == [8, 9] |
| 825 | render_request = ( |
| 826 | serving_responses_instance.online_renderer.render_responses.call_args.args[ |
| 827 | 0 |
| 828 | ] |
| 829 | ) |
| 830 | assert render_request.input == turn_messages |
| 831 | assert render_request.instructions is None |
| 832 | assert render_request.cache_salt == "request-salt" |
| 833 | |
| 834 | @pytest.mark.asyncio |
| 835 | async def test_harmony_tool_followup_preserves_render_params( |
| 836 | self, serving_responses_instance, monkeypatch: pytest.MonkeyPatch |
| 837 | ): |
| 838 | class ToolCallingHarmonyContext(HarmonyContext): |
| 839 | def __init__(self): |
| 840 | self._messages = [] |
| 841 | self.request = None |
| 842 | self._needs_tool_call = True |
| 843 | |
| 844 | def append_output(self, output) -> None: |
| 845 | pass |
| 846 | |
| 847 | def need_builtin_tool_call(self) -> bool: |
| 848 | return self._needs_tool_call |
| 849 | |
| 850 | async def call_tool(self): |
| 851 | self._needs_tool_call = False |
| 852 | return [] |
| 853 | |
| 854 | def append_tool_output(self, output) -> None: |
| 855 | self._messages.extend(output) |
| 856 | |
| 857 | async def generate_output(): |
| 858 | yield MagicMock() |
| 859 | |
| 860 | context = ToolCallingHarmonyContext() |
| 861 | serving_responses_instance.engine_client.generate.side_effect = ( |
| 862 | lambda *args, **kwargs: generate_output() |
| 863 | ) |
| 864 | monkeypatch.setattr( |
| 865 | "vllm.renderers.online_renderer.render_for_completion", |
| 866 | lambda messages: [1, 2, 3, 4], |
| 867 | ) |
| 868 | online_renderer = serving_responses_instance.online_renderer |
| 869 | online_renderer.renderer = SimpleNamespace( |
| 870 | tokenizer=SimpleNamespace(truncation_side="left") |
| 871 | ) |
| 872 | serving_responses_instance.online_renderer.render_responses_harmony_messages = ( |
| 873 | OnlineRenderer.render_responses_harmony_messages.__get__(online_renderer) |
| 874 | ) |
| 875 | serving_responses_instance._extract_prompt_len = MagicMock(return_value=2) |
| 876 | |
| 877 | async for _ in serving_responses_instance._generate_with_builtin_tools( |
| 878 | request_id="req", |
| 879 | engine_input=tokens_input([1], cache_salt="request-salt"), |
| 880 | sampling_params=MagicMock(), |
| 881 | context=context, |
| 882 | tok_params=TokenizeParams( |
| 883 | max_total_tokens=3, |
| 884 | max_output_tokens=1, |
| 885 | truncate_prompt_tokens=-1, |
| 886 | ), |
| 887 | ): |
| 888 | pass |
| 889 | |
| 890 | assert serving_responses_instance.engine_client.generate.call_count == 2 |
| 891 | assert context.request_metrics_cover_all_generation_turns is False |
| 892 | followup_engine_input = ( |
| 893 | serving_responses_instance.engine_client.generate.call_args_list[1].args[0] |
| 894 | ) |
| 895 | assert followup_engine_input.get("cache_salt") == "request-salt" |
| 896 | assert followup_engine_input["prompt_token_ids"] == [3, 4] |
| 897 | |
| 898 | @pytest.mark.asyncio |
| 899 | async def test_harmony_tool_followup_respects_max_output_tokens( |
| 900 | self, serving_responses_instance |
| 901 | ): |
| 902 | class ToolCallingHarmonyContext(HarmonyContext): |
| 903 | def __init__(self, request): |
| 904 | self._messages = [] |
| 905 | self.request = request |
| 906 | self._needs_tool_call = True |
| 907 | |
| 908 | def append_output(self, output) -> None: |
| 909 | pass |
| 910 | |
| 911 | def need_builtin_tool_call(self) -> bool: |
| 912 | return self._needs_tool_call |
| 913 | |
| 914 | async def call_tool(self): |
| 915 | self._needs_tool_call = False |
| 916 | return [] |
| 917 | |
| 918 | def append_tool_output(self, output) -> None: |
| 919 | pass |
| 920 | |
| 921 | max_tokens_per_turn = [] |
| 922 | |
| 923 | async def generate_output(engine_input, sampling_params, *args, **kwargs): |
| 924 | max_tokens_per_turn.append(sampling_params.max_tokens) |
| 925 | yield MagicMock() |
| 926 | |
| 927 | serving_responses_instance.engine_client.generate.side_effect = generate_output |
| 928 | serving_responses_instance.online_renderer.render_responses_harmony_messages = ( |
| 929 | MagicMock(return_value=tokens_input([1, 2, 3])) |
| 930 | ) |
| 931 | context = ToolCallingHarmonyContext( |
| 932 | ResponsesRequest(input="test", max_output_tokens=10) |
| 933 | ) |
| 934 | |
| 935 | async for _ in serving_responses_instance._generate_with_builtin_tools( |
| 936 | request_id="req", |
| 937 | engine_input=tokens_input([1]), |
| 938 | sampling_params=SamplingParams(max_tokens=10), |
| 939 | context=context, |
| 940 | ): |
| 941 | pass |
| 942 | |
| 943 | assert max_tokens_per_turn == [10, 10] |
| 944 | |
| 945 | @pytest.mark.asyncio |
| 946 | async def test_initialize_tool_sessions( |
| 947 | self, serving_responses_instance, mock_context, mock_exit_stack |
| 948 | ): |
| 949 | """Test that method works correctly with only MCP tools.""" |
| 950 | request = ResponsesRequest(input="test input", tools=[]) |
| 951 | |
| 952 | # Call the method |
| 953 | await serving_responses_instance._initialize_tool_sessions( |
| 954 | request, mock_context, mock_exit_stack |
| 955 | ) |
| 956 | assert mock_context.init_tool_sessions_called is False |
| 957 | |
| 958 | # Create only MCP tools |
| 959 | tools = [ |
| 960 | {"type": "web_search_preview"}, |
| 961 | {"type": "code_interpreter", "container": {"type": "auto"}}, |
| 962 | ] |
| 963 | |
| 964 | request = ResponsesRequest(input="test input", tools=tools) |
| 965 | |
| 966 | # Call the method |
| 967 | await serving_responses_instance._initialize_tool_sessions( |
| 968 | request, mock_context, mock_exit_stack |
| 969 | ) |
| 970 | |
| 971 | # Verify that init_tool_sessions was called |
| 972 | assert mock_context.init_tool_sessions_called |
| 973 | |
| 974 | def test_validate_create_responses_input( |
| 975 | self, serving_responses_instance, mock_context, mock_exit_stack |
| 976 | ): |
| 977 | request = ResponsesRequest( |
| 978 | input="test input", |
| 979 | previous_input_messages=[ |
| 980 | { |
| 981 | "role": "user", |
| 982 | "content": [ |
| 983 | { |
| 984 | "type": "text", |
| 985 | "text": "What is my horoscope? I am an Aquarius.", |
| 986 | } |
| 987 | ], |
| 988 | } |
| 989 | ], |
| 990 | previous_response_id="lol", |
| 991 | ) |
| 992 | error = serving_responses_instance._validate_create_responses_input(request) |
| 993 | assert error is not None |
| 994 | assert error.error.type == "invalid_request_error" |
| 995 | |
| 996 | |
| 997 | class TestValidateGeneratorInput: |
| 998 | """Test class for _validate_generator_input method.""" |
| 999 | |
| 1000 | @pytest_asyncio.fixture |
| 1001 | async def serving_responses_instance(self): |
| 1002 | """Create a real OpenAIServingResponses instance for testing.""" |
| 1003 | # Create minimal mocks for required dependencies |
| 1004 | engine_client = MagicMock() |
| 1005 | |
| 1006 | model_config = MagicMock() |
| 1007 | model_config.max_model_len = 100 |
| 1008 | model_config.hf_config.model_type = "test" |
| 1009 | model_config.get_diff_sampling_param.return_value = {} |
| 1010 | engine_client.model_config = model_config |
| 1011 | |
| 1012 | engine_client.input_processor = MagicMock() |
| 1013 | engine_client.renderer = MagicMock() |
| 1014 | |
| 1015 | models = MagicMock() |
| 1016 | |
| 1017 | # Create the actual instance |
| 1018 | instance = OpenAIServingResponses( |
| 1019 | engine_client=engine_client, |
| 1020 | models=models, |
| 1021 | online_renderer=MagicMock(), |
| 1022 | request_logger=None, |
| 1023 | chat_template=None, |
| 1024 | chat_template_content_format="auto", |
| 1025 | ) |
| 1026 | |
| 1027 | return instance |
| 1028 | |
| 1029 | def test_validate_generator_input(self, serving_responses_instance): |
| 1030 | """Test _validate_generator_input with valid prompt length.""" |
| 1031 | # Create an engine prompt with valid length (less than max_model_len) |
| 1032 | valid_prompt_token_ids = list(range(5)) # 5 tokens < 100 max_model_len |
| 1033 | engine_input = tokens_input(valid_prompt_token_ids) |
| 1034 | |
| 1035 | # Call the method |
| 1036 | result = serving_responses_instance._validate_generator_input(engine_input) |
| 1037 | |
| 1038 | # Should return None for valid input |
| 1039 | assert result is None |
| 1040 | |
| 1041 | # create an invalid engine prompt |
| 1042 | invalid_prompt_token_ids = list(range(200)) # 100 tokens >= 100 max_model_len |
| 1043 | engine_input = tokens_input(invalid_prompt_token_ids) |
| 1044 | |
| 1045 | # Call the method |
| 1046 | result = serving_responses_instance._validate_generator_input(engine_input) |
| 1047 | |
| 1048 | # Should return an ErrorResponse |
| 1049 | assert result is not None |
| 1050 | assert isinstance(result, ErrorResponse) |
| 1051 | |
| 1052 | |
| 1053 | class _Qwen3FakeTokenizer: |
| 1054 | def __init__(self): |
| 1055 | self._vocab = {"<think>": 1, "</think>": 2, "reason": 3, "final": 4} |
| 1056 | |
| 1057 | def get_vocab(self): |
| 1058 | return self._vocab |
| 1059 | |
| 1060 | def decode(self, token_ids): |
| 1061 | id_to_token = {v: k for k, v in self._vocab.items()} |
| 1062 | return "".join(id_to_token.get(token_id, "x") for token_id in token_ids) |
| 1063 | |
| 1064 | |
| 1065 | def _make_qwen3_serving() -> tuple[OpenAIServingResponses, _Qwen3FakeTokenizer]: |
| 1066 | engine_client = MagicMock() |
| 1067 | model_config = MagicMock() |
| 1068 | model_config.hf_config.model_type = "test" |
| 1069 | model_config.hf_text_config = MagicMock() |
| 1070 | model_config.get_diff_sampling_param.return_value = {} |
| 1071 | engine_client.model_config = model_config |
| 1072 | engine_client.input_processor = MagicMock() |
| 1073 | engine_client.renderer = MagicMock() |
| 1074 | |
| 1075 | tokenizer = _Qwen3FakeTokenizer() |
| 1076 | engine_client.renderer.get_tokenizer.return_value = tokenizer |
| 1077 | |
| 1078 | serving = OpenAIServingResponses( |
| 1079 | engine_client=engine_client, |
| 1080 | models=MagicMock(), |
| 1081 | online_renderer=MagicMock(), |
| 1082 | request_logger=None, |
| 1083 | chat_template=None, |
| 1084 | chat_template_content_format="auto", |
| 1085 | reasoning_parser="qwen3", |
| 1086 | ) |
| 1087 | return serving, tokenizer |
| 1088 | |
| 1089 | |
| 1090 | def _make_text_request_output(text: str, token_ids: list[int]) -> RequestOutput: |
| 1091 | return RequestOutput( |
| 1092 | request_id="req", |
| 1093 | prompt="hi", |
| 1094 | prompt_token_ids=[7, 8], |
| 1095 | prompt_logprobs=None, |
| 1096 | outputs=[ |
| 1097 | CompletionOutput( |
| 1098 | index=0, |
| 1099 | text=text, |
| 1100 | token_ids=token_ids, |
| 1101 | cumulative_logprob=0.0, |
| 1102 | logprobs=None, |
| 1103 | finish_reason="stop", |
| 1104 | stop_reason=None, |
| 1105 | ) |
| 1106 | ], |
| 1107 | finished=True, |
| 1108 | num_cached_tokens=0, |
| 1109 | num_cache_creation_tokens=3, |
| 1110 | ) |
| 1111 | |
| 1112 | |
| 1113 | async def _run_full_generator(serving, request, context, tokenizer): |
| 1114 | async def dummy_result_generator(): |
| 1115 | yield None |
| 1116 | |
| 1117 | return await serving.responses_full_generator( |
| 1118 | request=request, |
| 1119 | sampling_params=SamplingParams(max_tokens=16), |
| 1120 | result_generator=dummy_result_generator(), |
| 1121 | context=context, |
| 1122 | model_name="test-model", |
| 1123 | tokenizer=tokenizer, |
| 1124 | request_metadata=RequestResponseMetadata(request_id="req"), |
| 1125 | ) |
| 1126 | |
| 1127 | |
| 1128 | @pytest.mark.asyncio |
| 1129 | async def test_reasoning_tokens_counted_for_text_reasoning_model(monkeypatch): |
| 1130 | """Ensure reasoning_tokens usage is derived from thinking token spans.""" |
| 1131 | # Force non-harmony, SimpleContext path |
| 1132 | monkeypatch.setattr(envs, "VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", False) |
| 1133 | |
| 1134 | serving, tokenizer = _make_qwen3_serving() |
| 1135 | request = ResponsesRequest(input="hi", tools=[], stream=False) |
| 1136 | response_parser = serving._make_response_parser( |
| 1137 | request, |
| 1138 | tokenizer, |
| 1139 | serving._effective_chat_template_kwargs(request), |
| 1140 | ) |
| 1141 | |
| 1142 | # Build a SimpleContext with thinking tokens in the output. |
| 1143 | context = SimpleContext(response_parser=response_parser) |
| 1144 | # <think> 10 </think> 20 -> reasoning token count = 1 |
| 1145 | context.append_output( |
| 1146 | _make_text_request_output("<think>reason</think>final", [1, 10, 2, 20]) |
| 1147 | ) |
| 1148 | |
| 1149 | response = await _run_full_generator(serving, request, context, tokenizer) |
| 1150 | |
| 1151 | assert response.usage.output_tokens_details.reasoning_tokens == 1 |
| 1152 | assert response.usage.input_tokens_details.cache_write_tokens == 3 |
| 1153 | |
| 1154 | OpenAIResponseUsage.model_validate(response.usage.model_dump(mode="json")) |
| 1155 | |
| 1156 | |
| 1157 | class TestExtractAllowedToolsFromMcpRequests: |
| 1158 | """Test class for _extract_allowed_tools_from_mcp_requests function.""" |
| 1159 | |
| 1160 | def test_extract_allowed_tools_basic_formats(self): |
| 1161 | """Test extraction with list format, object format, and None.""" |
| 1162 | from openai.types.responses.tool import McpAllowedToolsMcpToolFilter |
| 1163 | |
| 1164 | tools = [ |
| 1165 | # List format |
| 1166 | Mcp( |
| 1167 | type="mcp", |
| 1168 | server_label="server1", |
| 1169 | allowed_tools=["tool1", "tool2"], |
| 1170 | ), |
| 1171 | # Object format |
| 1172 | Mcp( |
| 1173 | type="mcp", |
| 1174 | server_label="server2", |
| 1175 | allowed_tools=McpAllowedToolsMcpToolFilter( |
| 1176 | tool_names=["tool3", "tool4"] |
| 1177 | ), |
| 1178 | ), |
| 1179 | # None (no filter) |
| 1180 | Mcp( |
| 1181 | type="mcp", |
| 1182 | server_label="server3", |
| 1183 | allowed_tools=None, |
| 1184 | ), |
| 1185 | ] |
| 1186 | result = _extract_allowed_tools_from_mcp_requests(tools) |
| 1187 | assert result == { |
| 1188 | "server1": ["tool1", "tool2"], |
| 1189 | "server2": ["tool3", "tool4"], |
| 1190 | "server3": None, |
| 1191 | } |
| 1192 | |
| 1193 | def test_extract_allowed_tools_star_normalization(self): |
| 1194 | """Test that '*' wildcard is normalized to None (select all tools). |
| 1195 | |
| 1196 | This is the key test requested by reviewers to explicitly demonstrate |
| 1197 | that the "*" select-all scenario is handled correctly. |
| 1198 | """ |
| 1199 | from openai.types.responses.tool import McpAllowedToolsMcpToolFilter |
| 1200 | |
| 1201 | tools = [ |
| 1202 | # Star in list format |
| 1203 | Mcp( |
| 1204 | type="mcp", |
| 1205 | server_label="server1", |
| 1206 | allowed_tools=["*"], |
| 1207 | ), |
| 1208 | # Star mixed with other tools in list |
| 1209 | Mcp( |
| 1210 | type="mcp", |
| 1211 | server_label="server2", |
| 1212 | allowed_tools=["tool1", "*"], |
| 1213 | ), |
| 1214 | # Star in object format |
| 1215 | Mcp( |
| 1216 | type="mcp", |
| 1217 | server_label="server3", |
| 1218 | allowed_tools=McpAllowedToolsMcpToolFilter(tool_names=["*"]), |
| 1219 | ), |
| 1220 | ] |
| 1221 | result = _extract_allowed_tools_from_mcp_requests(tools) |
| 1222 | # All should be normalized to None (allows all tools) |
| 1223 | assert result == { |
| 1224 | "server1": None, |
| 1225 | "server2": None, |
| 1226 | "server3": None, |
| 1227 | } |
| 1228 | |
| 1229 | def test_extract_allowed_tools_filters_non_mcp(self): |
| 1230 | """Test that non-MCP tools are ignored during extraction.""" |
| 1231 | tools = [ |
| 1232 | Mcp( |
| 1233 | type="mcp", |
| 1234 | server_label="server1", |
| 1235 | allowed_tools=["tool1"], |
| 1236 | ), |
| 1237 | LocalShell(type="local_shell"), # Non-MCP tool should be ignored |
| 1238 | Mcp( |
| 1239 | type="mcp", |
| 1240 | server_label="server2", |
| 1241 | allowed_tools=["tool2"], |
| 1242 | ), |
| 1243 | ] |
| 1244 | result = _extract_allowed_tools_from_mcp_requests(tools) |
| 1245 | # Non-MCP tools should be ignored |
| 1246 | assert result == { |
| 1247 | "server1": ["tool1"], |
| 1248 | "server2": ["tool2"], |
| 1249 | } |
| 1250 | |
| 1251 | |
| 1252 | class TestHarmonyPreambleStreaming: |
| 1253 | """Tests for preamble (commentary with no recipient) streaming events.""" |
| 1254 | |
| 1255 | @staticmethod |
| 1256 | def _make_segment(*, channel, recipient, delta="hello"): |
| 1257 | """Build a lightweight segment for Harmony streaming tests.""" |
| 1258 | return Segment(channel=channel, recipient=recipient, delta=delta) |
| 1259 | |
| 1260 | @staticmethod |
| 1261 | def _make_previous_item(*, channel, recipient, text="preamble text"): |
| 1262 | """Build a lightweight mock previous_item (openai_harmony Message).""" |
| 1263 | content_part = MagicMock() |
| 1264 | content_part.text = text |
| 1265 | item = MagicMock() |
| 1266 | item.channel = channel |
| 1267 | item.recipient = recipient |
| 1268 | item.content = [content_part] |
| 1269 | return item |
| 1270 | |
| 1271 | def test_preamble_delta_emits_text_events(self) -> None: |
| 1272 | """Commentary + recipient=None should emit output_text.delta events.""" |
| 1273 | from vllm.entrypoints.openai.responses.streaming_events import ( |
| 1274 | emit_content_delta_events, |
| 1275 | ) |
| 1276 | |
| 1277 | segment = self._make_segment(channel="commentary", recipient=None) |
| 1278 | state = StreamingState() |
| 1279 | |
| 1280 | events = emit_content_delta_events(segment, state) |
| 1281 | |
| 1282 | type_names = [e.type for e in events] |
| 1283 | assert "response.output_text.delta" in type_names |
| 1284 | assert "response.output_item.added" in type_names |
| 1285 | |
| 1286 | def test_preamble_delta_second_token_no_added(self) -> None: |
| 1287 | """Second preamble token should emit delta only, not added again.""" |
| 1288 | from vllm.entrypoints.openai.responses.streaming_events import ( |
| 1289 | emit_content_delta_events, |
| 1290 | ) |
| 1291 | |
| 1292 | segment = self._make_segment(channel="commentary", recipient=None, delta="w") |
| 1293 | state = StreamingState() |
| 1294 | state.sent_output_item_added = True |
| 1295 | state.current_item_id = "msg_test" |
| 1296 | state.current_content_index = 0 |
| 1297 | |
| 1298 | events = emit_content_delta_events(segment, state) |
| 1299 | |
| 1300 | type_names = [e.type for e in events] |
| 1301 | assert "response.output_text.delta" in type_names |
| 1302 | assert "response.output_item.added" not in type_names |
| 1303 | |
| 1304 | def test_commentary_with_function_recipient_not_preamble(self) -> None: |
| 1305 | """Commentary + recipient='functions.X' must NOT use preamble path.""" |
| 1306 | from vllm.entrypoints.openai.responses.streaming_events import ( |
| 1307 | emit_content_delta_events, |
| 1308 | ) |
| 1309 | |
| 1310 | segment = self._make_segment( |
| 1311 | channel="commentary", |
| 1312 | recipient="functions.get_weather", |
| 1313 | ) |
| 1314 | state = StreamingState() |
| 1315 | |
| 1316 | events = emit_content_delta_events(segment, state) |
| 1317 | |
| 1318 | type_names = [e.type for e in events] |
| 1319 | assert "response.output_text.delta" not in type_names |
| 1320 | |
| 1321 | def test_preamble_done_emits_text_done_events(self) -> None: |
| 1322 | """Completed preamble should emit text done + content_part done + |
| 1323 | output_item done, same shape as final channel.""" |
| 1324 | from vllm.entrypoints.openai.responses.streaming_events import ( |
| 1325 | emit_previous_item_done_events, |
| 1326 | ) |
| 1327 | |
| 1328 | previous = self._make_previous_item(channel="commentary", recipient=None) |
| 1329 | state = StreamingState() |
| 1330 | state.sent_output_item_added = True |
| 1331 | state.current_item_id = "msg_test" |
| 1332 | state.current_output_index = 0 |
| 1333 | state.current_content_index = 0 |
| 1334 | |
| 1335 | events = emit_previous_item_done_events(previous, state) |
| 1336 | |
| 1337 | type_names = [e.type for e in events] |
| 1338 | assert "response.output_text.done" in type_names |
| 1339 | assert "response.content_part.done" in type_names |
| 1340 | assert "response.output_item.done" in type_names |
| 1341 | |
| 1342 | def test_commentary_with_recipient_no_preamble_done(self) -> None: |
| 1343 | """Commentary + recipient='functions.X' should route to function call |
| 1344 | done, not preamble done.""" |
| 1345 | from vllm.entrypoints.openai.responses.streaming_events import ( |
| 1346 | emit_previous_item_done_events, |
| 1347 | ) |
| 1348 | |
| 1349 | previous = self._make_previous_item( |
| 1350 | channel="commentary", recipient="functions.get_weather" |
| 1351 | ) |
| 1352 | state = StreamingState() |
| 1353 | state.is_first_function_call_delta = True |
| 1354 | state.current_item_id = "fc_test" |
| 1355 | state.current_call_id = "call_test" |
| 1356 | |
| 1357 | events = emit_previous_item_done_events(previous, state) |
| 1358 | |
| 1359 | type_names = [e.type for e in events] |
| 1360 | assert "response.output_text.done" not in type_names |
| 1361 | |
| 1362 | @pytest.mark.xfail( |
| 1363 | reason=( |
| 1364 | "TODO: Ensure added/in-progress events are emitted for zero-delta items." |
| 1365 | "So we can safely emit done events for zero-delta items." |
| 1366 | ), |
| 1367 | strict=True, |
| 1368 | ) |
| 1369 | def test_zero_delta_items_should_preserve_streaming_lifecycle( |
| 1370 | self, |
| 1371 | ) -> None: |
| 1372 | """Zero-delta Harmony items should still produce a coherent lifecycle.""" |
| 1373 | from vllm.entrypoints.openai.responses.streaming_events import ( |
| 1374 | emit_previous_item_done_events, |
| 1375 | ) |
| 1376 | |
| 1377 | cases: list[tuple[str, str | None, str]] = [ |
| 1378 | ("commentary", None, "msg_stale"), |
| 1379 | ("analysis", None, "msg_stale"), |
| 1380 | ("commentary", "functions.get_weather", "fc_stale"), |
| 1381 | ("commentary", "python", "tool_stale"), |
| 1382 | ("commentary", "repo_browser.list", "mcp_stale"), |
| 1383 | ] |
| 1384 | |
| 1385 | for channel, recipient, current_item_id in cases: |
| 1386 | previous = self._make_previous_item(channel=channel, recipient=recipient) |
| 1387 | state = StreamingState() |
| 1388 | state.current_item_id = current_item_id |
| 1389 | state.current_call_id = "call_stale" |
| 1390 | state.current_content_index = 0 |
| 1391 | |
| 1392 | events = emit_previous_item_done_events( |
| 1393 | previous, state, function_tool_names=None |
| 1394 | ) |
| 1395 | |
| 1396 | type_names = [e.type for e in events] |
| 1397 | assert "response.output_item.added" in type_names |
| 1398 | assert "response.output_item.done" in type_names |
| 1399 | |
| 1400 | |
| 1401 | _PER_REQUEST_STATS = RequestStateStats( |
| 1402 | queued_ts=1.0, |
| 1403 | scheduled_ts=1.5, |
| 1404 | first_token_ts=2.0, |
| 1405 | last_token_ts=3.0, |
| 1406 | ) |
| 1407 | |
| 1408 | |
| 1409 | def _make_request_output( |
| 1410 | text, |
| 1411 | token_ids, |
| 1412 | metrics: RequestStateStats | None = None, |
| 1413 | ): |
| 1414 | completion = CompletionOutput( |
| 1415 | index=0, |
| 1416 | text=text, |
| 1417 | token_ids=token_ids, |
| 1418 | cumulative_logprob=0.0, |
| 1419 | logprobs=None, |
| 1420 | finish_reason=None, |
| 1421 | stop_reason=None, |
| 1422 | ) |
| 1423 | return RequestOutput( |
| 1424 | request_id="req", |
| 1425 | prompt="hi", |
| 1426 | prompt_token_ids=[7, 8], |
| 1427 | prompt_logprobs=None, |
| 1428 | outputs=[completion], |
| 1429 | finished=False, |
| 1430 | num_cached_tokens=0, |
| 1431 | metrics=metrics, |
| 1432 | ) |
| 1433 | |
| 1434 | |
| 1435 | def _make_simple_context_with_output( |
| 1436 | text, |
| 1437 | token_ids, |
| 1438 | response_parser=None, |
| 1439 | metrics: RequestStateStats | None = None, |
| 1440 | ): |
| 1441 | """Create a SimpleContext with a RequestOutput containing the given text.""" |
| 1442 | ctx = SimpleContext(response_parser=response_parser) |
| 1443 | req_output = _make_request_output(text, token_ids, metrics) |
| 1444 | ctx.append_output(req_output) |
| 1445 | ctx.request_metrics = metrics |
| 1446 | return ctx |
| 1447 | |
| 1448 | |
| 1449 | def _make_serving_instance( |
| 1450 | *, |
| 1451 | reasoning_parser: str = "", |
| 1452 | tool_parser: str | None = None, |
| 1453 | enable_per_request_metrics: bool = False, |
| 1454 | ) -> OpenAIServingResponses: |
| 1455 | engine_client = MagicMock() |
| 1456 | model_config = MagicMock() |
| 1457 | model_config.max_model_len = 100 |
| 1458 | model_config.hf_config.model_type = "test" |
| 1459 | model_config.hf_text_config = MagicMock() |
| 1460 | model_config.get_diff_sampling_param.return_value = {} |
| 1461 | engine_client.model_config = model_config |
| 1462 | engine_client.input_processor = MagicMock() |
| 1463 | engine_client.renderer = MagicMock() |
| 1464 | |
| 1465 | return OpenAIServingResponses( |
| 1466 | engine_client=engine_client, |
| 1467 | models=MagicMock(), |
| 1468 | online_renderer=MagicMock(), |
| 1469 | request_logger=None, |
| 1470 | chat_template=None, |
| 1471 | chat_template_content_format="auto", |
| 1472 | reasoning_parser=reasoning_parser, |
| 1473 | enable_auto_tools=tool_parser is not None, |
| 1474 | tool_parser=tool_parser, |
| 1475 | enable_per_request_metrics=enable_per_request_metrics, |
| 1476 | ) |
| 1477 | |
| 1478 | |
| 1479 | async def _empty_context_generator(): |
| 1480 | if False: |
| 1481 | yield |
| 1482 | |
| 1483 | |
| 1484 | async def _make_full_metrics_response( |
| 1485 | enable_per_request_metrics: bool, |
| 1486 | request_metrics_cover_all_generation_turns: bool = True, |
| 1487 | ): |
| 1488 | serving = _make_serving_instance( |
| 1489 | enable_per_request_metrics=enable_per_request_metrics |
| 1490 | ) |
| 1491 | request = ResponsesRequest(input="hi", tools=[], stream=False, store=False) |
| 1492 | sampling_params = SamplingParams(max_tokens=16) |
| 1493 | context = SimpleContext() |
| 1494 | context.request_metrics_cover_all_generation_turns = ( |
| 1495 | request_metrics_cover_all_generation_turns |
| 1496 | ) |
| 1497 | |
| 1498 | async def generate(*args, **kwargs): |
| 1499 | yield _make_request_output("hello", [10, 20], _PER_REQUEST_STATS) |
| 1500 | |
| 1501 | serving.engine_client.generate.side_effect = generate |
| 1502 | result_generator = serving._generate_with_builtin_tools( |
| 1503 | request_id=request.request_id, |
| 1504 | engine_input=tokens_input([7, 8]), |
| 1505 | sampling_params=sampling_params, |
| 1506 | context=context, |
| 1507 | ) |
| 1508 | response = await serving.responses_full_generator( |
| 1509 | request=request, |
| 1510 | sampling_params=sampling_params, |
| 1511 | result_generator=result_generator, |
| 1512 | context=context, |
| 1513 | model_name="test-model", |
| 1514 | tokenizer=MagicMock(), |
| 1515 | request_metadata=RequestResponseMetadata(request_id="req"), |
| 1516 | ) |
| 1517 | assert isinstance(response, ResponsesResponse) |
| 1518 | return response |
| 1519 | |
| 1520 | |
| 1521 | @pytest.mark.asyncio |
| 1522 | async def test_responses_per_request_metrics_follow_server_flag(): |
| 1523 | disabled_response = await _make_full_metrics_response(False) |
| 1524 | assert disabled_response.metrics is None |
| 1525 | assert "metrics" not in disabled_response.model_dump(mode="json") |
| 1526 | |
| 1527 | enabled_response = await _make_full_metrics_response(True) |
| 1528 | assert enabled_response.metrics is not None |
| 1529 | assert enabled_response.metrics.time_to_first_token_ms == pytest.approx(500.0) |
| 1530 | |
| 1531 | |
| 1532 | @pytest.mark.asyncio |
| 1533 | async def test_responses_streaming_metrics_only_on_completed_event(): |
| 1534 | serving = _make_serving_instance(enable_per_request_metrics=True) |
| 1535 | request = ResponsesRequest(input="hi", tools=[], stream=True, store=False) |
| 1536 | context = _make_simple_context_with_output( |
| 1537 | "hello", [10, 20], metrics=_PER_REQUEST_STATS |
| 1538 | ) |
| 1539 | |
| 1540 | events = [ |
| 1541 | event |
| 1542 | async for event in serving.responses_stream_generator( |
| 1543 | request=request, |
| 1544 | sampling_params=SamplingParams(max_tokens=16), |
| 1545 | result_generator=_empty_context_generator(), |
| 1546 | context=context, |
| 1547 | model_name="test-model", |
| 1548 | tokenizer=MagicMock(), |
| 1549 | request_metadata=RequestResponseMetadata(request_id="req"), |
| 1550 | ) |
| 1551 | ] |
| 1552 | |
| 1553 | for event in events[:-1]: |
| 1554 | assert "metrics" not in event.response.model_dump(mode="json") |
| 1555 | assert isinstance(events[-1], ResponseCompletedEvent) |
| 1556 | assert events[-1].response.metrics is not None |
| 1557 | assert events[-1].response.metrics.time_to_first_token_ms == pytest.approx(500.0) |
| 1558 | |
| 1559 | |
| 1560 | @pytest.mark.asyncio |
| 1561 | async def test_responses_metrics_suppressed_for_multiple_generation_turns(): |
| 1562 | response = await _make_full_metrics_response( |
| 1563 | True, request_metrics_cover_all_generation_turns=False |
| 1564 | ) |
| 1565 | |
| 1566 | assert response.metrics is None |
| 1567 | assert "metrics" not in response.model_dump(mode="json") |
| 1568 | |
| 1569 | |
| 1570 | def _make_serving_instance_with_reasoning(): |
| 1571 | """Create an OpenAIServingResponses with a mocked reasoning parser.""" |
| 1572 | return _make_serving_instance(reasoning_parser="qwen3") |
| 1573 | |
| 1574 | |
| 1575 | def _identity_increment(event): |
| 1576 | """Simple identity callable for _increment_sequence_number_and_return.""" |
| 1577 | seq = getattr(_identity_increment, "_counter", 0) |
| 1578 | if hasattr(event, "sequence_number"): |
| 1579 | event.sequence_number = seq |
| 1580 | _identity_increment._counter = seq + 1 # type: ignore |
| 1581 | return event |
| 1582 | |
| 1583 | |
| 1584 | def _mock_parser_with_reasoning(serving, delta_sequence: list[DeltaMessage]): |
| 1585 | """Set up serving.parser so that it returns a mock parser instance |
| 1586 | with a reasoning parser that returns the given delta_sequence. |
| 1587 | |
| 1588 | The mock has reasoning_parser set (truthy) but tool_parser as None, |
| 1589 | so the parser's parse_delta enters the reasoning-only branch. |
| 1590 | """ |
| 1591 | call_count = 0 |
| 1592 | |
| 1593 | def mock_parse_delta(**kwargs): |
| 1594 | nonlocal call_count |
| 1595 | if call_count >= len(delta_sequence): |
| 1596 | return None |
| 1597 | result = delta_sequence[call_count] |
| 1598 | call_count += 1 |
| 1599 | return result |
| 1600 | |
| 1601 | mock_parser_instance = MagicMock() |
| 1602 | mock_parser_instance.reasoning_parser = MagicMock() # truthy |
| 1603 | mock_parser_instance.tool_parser = None |
| 1604 | mock_parser_instance.parse_delta = mock_parse_delta |
| 1605 | mock_parser_instance.is_reasoning_end = MagicMock(return_value=False) |
| 1606 | serving.parser = MagicMock(return_value=mock_parser_instance) |
| 1607 | return mock_parser_instance |
| 1608 | |
| 1609 | |
| 1610 | class TestStreamingReasoningToContentTransition: |
| 1611 | """Tests for _process_simple_streaming_events reasoning-to-content |
| 1612 | transition, specifically the fix for mixed deltas that carry both |
| 1613 | reasoning and content simultaneously.""" |
| 1614 | |
| 1615 | @pytest.mark.asyncio |
| 1616 | async def test_mixed_delta_reasoning_and_content_emits_reasoning_delta( |
| 1617 | self, monkeypatch |
| 1618 | ): |
| 1619 | """When the reasoning parser produces a delta with both reasoning |
| 1620 | and content set (e.g. reasoning end and content start in the same |
| 1621 | chunk), the trailing reasoning text must be emitted as a |
| 1622 | ResponseReasoningTextDeltaEvent and included in the |
| 1623 | ResponseReasoningTextDoneEvent text.""" |
| 1624 | monkeypatch.setattr(envs, "VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", False) |
| 1625 | serving = _make_serving_instance_with_reasoning() |
| 1626 | |
| 1627 | # Sequence of DeltaMessages the mock orchestrator will return |
| 1628 | delta_sequence = [ |
| 1629 | DeltaMessage(reasoning="thinking..."), |
| 1630 | DeltaMessage(reasoning=" end", content="hello"), # mixed delta |
| 1631 | DeltaMessage(content=" world"), |
| 1632 | ] |
| 1633 | response_parser = _mock_parser_with_reasoning(serving, delta_sequence) |
| 1634 | # Create contexts for each streaming chunk |
| 1635 | contexts = [ |
| 1636 | _make_simple_context_with_output("chunk1", [10], response_parser), |
| 1637 | _make_simple_context_with_output("chunk2", [20], response_parser), |
| 1638 | _make_simple_context_with_output("chunk3", [30], response_parser), |
| 1639 | ] |
| 1640 | |
| 1641 | async def result_generator(): |
| 1642 | for ctx in contexts: |
| 1643 | yield ctx |
| 1644 | |
| 1645 | request = ResponsesRequest(input="hi", tools=[], stream=True) |
| 1646 | sampling_params = SamplingParams(max_tokens=64) |
| 1647 | metadata = RequestResponseMetadata(request_id="req") |
| 1648 | _identity_increment._counter = 0 # type: ignore |
| 1649 | |
| 1650 | events = [] |
| 1651 | async for event in serving._process_simple_streaming_events( |
| 1652 | request=request, |
| 1653 | sampling_params=sampling_params, |
| 1654 | result_generator=result_generator(), |
| 1655 | context=SimpleContext(response_parser=response_parser), |
| 1656 | model_name="test-model", |
| 1657 | tokenizer=MagicMock(), |
| 1658 | request_metadata=metadata, |
| 1659 | created_time=0, |
| 1660 | _increment_sequence_number_and_return=_identity_increment, |
| 1661 | ): |
| 1662 | events.append(event) |
| 1663 | |
| 1664 | # The first reasoning delta should be emitted |
| 1665 | reasoning_deltas = [ |
| 1666 | e for e in events if isinstance(e, ResponseReasoningTextDeltaEvent) |
| 1667 | ] |
| 1668 | assert len(reasoning_deltas) == 2 |
| 1669 | assert reasoning_deltas[0].delta == "thinking..." |
| 1670 | # The trailing reasoning from the mixed delta must also be emitted |
| 1671 | assert reasoning_deltas[1].delta == " end" |
| 1672 | |
| 1673 | # The done event must include both reasoning parts |
| 1674 | reasoning_done = [ |
| 1675 | e for e in events if isinstance(e, ResponseReasoningTextDoneEvent) |
| 1676 | ] |
| 1677 | assert len(reasoning_done) == 1 |
| 1678 | assert reasoning_done[0].text == "thinking... end" |
| 1679 | |
| 1680 | # Content deltas should be emitted for both the mixed delta's |
| 1681 | # content and the pure content delta |
| 1682 | text_deltas = [e for e in events if isinstance(e, ResponseTextDeltaEvent)] |
| 1683 | assert len(text_deltas) == 2 |
| 1684 | assert text_deltas[0].delta == "hello" |
| 1685 | assert text_deltas[1].delta == " world" |
| 1686 | |
| 1687 | @pytest.mark.asyncio |
| 1688 | async def test_transition_without_mixed_delta_no_extra_reasoning_event( |
| 1689 | self, monkeypatch |
| 1690 | ): |
| 1691 | """When the transition from reasoning to content is clean (no mixed |
| 1692 | delta), no extra reasoning delta event should be emitted.""" |
| 1693 | monkeypatch.setattr(envs, "VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", False) |
| 1694 | serving = _make_serving_instance_with_reasoning() |
| 1695 | |
| 1696 | delta_sequence = [ |
| 1697 | DeltaMessage(reasoning="thinking"), |
| 1698 | DeltaMessage(content="answer"), |
| 1699 | ] |
| 1700 | response_parser = _mock_parser_with_reasoning(serving, delta_sequence) |
| 1701 | |
| 1702 | contexts = [ |
| 1703 | _make_simple_context_with_output("chunk1", [10], response_parser), |
| 1704 | _make_simple_context_with_output("chunk2", [20], response_parser), |
| 1705 | ] |
| 1706 | |
| 1707 | async def result_generator(): |
| 1708 | for ctx in contexts: |
| 1709 | yield ctx |
| 1710 | |
| 1711 | request = ResponsesRequest(input="hi", tools=[], stream=True) |
| 1712 | sampling_params = SamplingParams(max_tokens=64) |
| 1713 | metadata = RequestResponseMetadata(request_id="req") |
| 1714 | _identity_increment._counter = 0 # type: ignore |
| 1715 | |
| 1716 | events = [] |
| 1717 | async for event in serving._process_simple_streaming_events( |
| 1718 | request=request, |
| 1719 | sampling_params=sampling_params, |
| 1720 | result_generator=result_generator(), |
| 1721 | context=SimpleContext(response_parser=response_parser), |
| 1722 | model_name="test-model", |
| 1723 | tokenizer=MagicMock(), |
| 1724 | request_metadata=metadata, |
| 1725 | created_time=0, |
| 1726 | _increment_sequence_number_and_return=_identity_increment, |
| 1727 | ): |
| 1728 | events.append(event) |
| 1729 | |
| 1730 | # Exactly one reasoning delta |
| 1731 | reasoning_deltas = [ |
| 1732 | e for e in events if isinstance(e, ResponseReasoningTextDeltaEvent) |
| 1733 | ] |
| 1734 | assert len(reasoning_deltas) == 1 |
| 1735 | assert reasoning_deltas[0].delta == "thinking" |
| 1736 | |
| 1737 | # Done event has just "thinking" |
| 1738 | reasoning_done = [ |
| 1739 | e for e in events if isinstance(e, ResponseReasoningTextDoneEvent) |
| 1740 | ] |
| 1741 | assert len(reasoning_done) == 1 |
| 1742 | assert reasoning_done[0].text == "thinking" |
| 1743 | |
| 1744 | # One content delta |
| 1745 | text_deltas = [e for e in events if isinstance(e, ResponseTextDeltaEvent)] |
| 1746 | assert len(text_deltas) == 1 |
| 1747 | assert text_deltas[0].delta == "answer" |
| 1748 | |
| 1749 | @pytest.mark.asyncio |
| 1750 | async def test_reasoning_only_stream_no_content(self, monkeypatch): |
| 1751 | """When the stream has only reasoning deltas and no content, the |
| 1752 | reasoning done event should be emitted at finalization with the |
| 1753 | full accumulated text, and no text delta events should appear.""" |
| 1754 | monkeypatch.setattr(envs, "VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", False) |
| 1755 | serving = _make_serving_instance_with_reasoning() |
| 1756 | |
| 1757 | delta_sequence = [ |
| 1758 | DeltaMessage(reasoning="step 1"), |
| 1759 | DeltaMessage(reasoning=" step 2"), |
| 1760 | ] |
| 1761 | response_parser = _mock_parser_with_reasoning(serving, delta_sequence) |
| 1762 | |
| 1763 | contexts = [ |
| 1764 | _make_simple_context_with_output("chunk1", [10], response_parser), |
| 1765 | _make_simple_context_with_output("chunk2", [20], response_parser), |
| 1766 | ] |
| 1767 | |
| 1768 | async def result_generator(): |
| 1769 | for ctx in contexts: |
| 1770 | yield ctx |
| 1771 | |
| 1772 | request = ResponsesRequest(input="hi", tools=[], stream=True) |
| 1773 | sampling_params = SamplingParams(max_tokens=64) |
| 1774 | metadata = RequestResponseMetadata(request_id="req") |
| 1775 | _identity_increment._counter = 0 # type: ignore |
| 1776 | |
| 1777 | events = [] |
| 1778 | async for event in serving._process_simple_streaming_events( |
| 1779 | request=request, |
| 1780 | sampling_params=sampling_params, |
| 1781 | result_generator=result_generator(), |
| 1782 | context=SimpleContext(response_parser=response_parser), |
| 1783 | model_name="test-model", |
| 1784 | tokenizer=MagicMock(), |
| 1785 | request_metadata=metadata, |
| 1786 | created_time=0, |
| 1787 | _increment_sequence_number_and_return=_identity_increment, |
| 1788 | ): |
| 1789 | events.append(event) |
| 1790 | |
| 1791 | # Two reasoning deltas |
| 1792 | reasoning_deltas = [ |
| 1793 | e for e in events if isinstance(e, ResponseReasoningTextDeltaEvent) |
| 1794 | ] |
| 1795 | assert len(reasoning_deltas) == 2 |
| 1796 | assert reasoning_deltas[0].delta == "step 1" |
| 1797 | assert reasoning_deltas[1].delta == " step 2" |
| 1798 | |
| 1799 | # Done event at finalization with accumulated text |
| 1800 | reasoning_done = [ |
| 1801 | e for e in events if isinstance(e, ResponseReasoningTextDoneEvent) |
| 1802 | ] |
| 1803 | assert len(reasoning_done) == 1 |
| 1804 | assert reasoning_done[0].text == "step 1 step 2" |
| 1805 | |
| 1806 | # No content text deltas |
| 1807 | text_deltas = [e for e in events if isinstance(e, ResponseTextDeltaEvent)] |
| 1808 | assert len(text_deltas) == 0 |
| 1809 | |
| 1810 | # Final item should be a reasoning item |
| 1811 | item_done_events = [ |
| 1812 | e for e in events if isinstance(e, ResponseOutputItemDoneEvent) |
| 1813 | ] |
| 1814 | assert len(item_done_events) == 1 |
| 1815 | assert isinstance(item_done_events[0].item, ResponseReasoningItem) |
| 1816 | |
| 1817 | |
| 1818 | class TestAutoToolStreaming: |
| 1819 | @staticmethod |
| 1820 | async def _collect_events(delta_sequence: list[DeltaMessage]): |
| 1821 | serving = _make_serving_instance_with_reasoning() |
| 1822 | response_parser = _mock_parser_with_reasoning(serving, delta_sequence) |
| 1823 | |
| 1824 | contexts = [ |
| 1825 | _make_simple_context_with_output("chunk", [i], response_parser) |
| 1826 | for i in range(len(delta_sequence)) |
| 1827 | ] |
| 1828 | |
| 1829 | async def result_generator(): |
| 1830 | for ctx in contexts: |
| 1831 | yield ctx |
| 1832 | |
| 1833 | request = ResponsesRequest( |
| 1834 | input="hi", |
| 1835 | tools=[ |
| 1836 | { |
| 1837 | "type": "function", |
| 1838 | "name": "get_weather", |
| 1839 | "description": "Get weather.", |
| 1840 | "parameters": { |
| 1841 | "type": "object", |
| 1842 | "properties": {"location": {"type": "string"}}, |
| 1843 | "required": ["location"], |
| 1844 | "additionalProperties": False, |
| 1845 | }, |
| 1846 | } |
| 1847 | ], |
| 1848 | tool_choice="auto", |
| 1849 | stream=True, |
| 1850 | ) |
| 1851 | sampling_params = SamplingParams(max_tokens=64) |
| 1852 | metadata = RequestResponseMetadata(request_id="req") |
| 1853 | _identity_increment._counter = 0 # type: ignore |
| 1854 | |
| 1855 | events = [] |
| 1856 | async for event in serving._process_simple_streaming_events( |
| 1857 | request=request, |
| 1858 | sampling_params=sampling_params, |
| 1859 | result_generator=result_generator(), |
| 1860 | context=SimpleContext(response_parser=response_parser), |
| 1861 | model_name="test-model", |
| 1862 | tokenizer=MagicMock(), |
| 1863 | request_metadata=metadata, |
| 1864 | created_time=0, |
| 1865 | _increment_sequence_number_and_return=_identity_increment, |
| 1866 | ): |
| 1867 | events.append(event) |
| 1868 | return events |
| 1869 | |
| 1870 | @pytest.mark.skip_global_cleanup |
| 1871 | @pytest.mark.asyncio |
| 1872 | async def test_auto_multi_tool_streaming_opens_one_item_per_tool(self, monkeypatch): |
| 1873 | monkeypatch.setattr(envs, "VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", False) |
| 1874 | |
| 1875 | delta_sequence = [ |
| 1876 | DeltaMessage( |
| 1877 | tool_calls=[ |
| 1878 | DeltaToolCall( |
| 1879 | id="call_vienna", |
| 1880 | type="function", |
| 1881 | index=0, |
| 1882 | function=DeltaFunctionCall( |
| 1883 | name="get_weather", |
| 1884 | arguments="", |
| 1885 | ), |
| 1886 | ) |
| 1887 | ] |
| 1888 | ), |
| 1889 | DeltaMessage( |
| 1890 | tool_calls=[ |
| 1891 | DeltaToolCall( |
| 1892 | index=0, |
| 1893 | function=DeltaFunctionCall( |
| 1894 | arguments='{"location":"Vienna"}', |
| 1895 | ), |
| 1896 | ) |
| 1897 | ] |
| 1898 | ), |
| 1899 | DeltaMessage( |
| 1900 | tool_calls=[ |
| 1901 | DeltaToolCall( |
| 1902 | id="call_berlin", |
| 1903 | type="function", |
| 1904 | index=1, |
| 1905 | function=DeltaFunctionCall( |
| 1906 | name="get_weather", |
| 1907 | arguments='{"location":"Berlin"}', |
| 1908 | ), |
| 1909 | ) |
| 1910 | ] |
| 1911 | ), |
| 1912 | ] |
| 1913 | events = await self._collect_events(delta_sequence) |
| 1914 | |
| 1915 | function_items = [ |
| 1916 | event |
| 1917 | for event in events |
| 1918 | if event.type == "response.output_item.added" |
| 1919 | and getattr(event.item, "type", None) == "function_call" |
| 1920 | ] |
| 1921 | assert len(function_items) == 2 |
| 1922 | assert [event.item.name for event in function_items] == [ |
| 1923 | "get_weather", |
| 1924 | "get_weather", |
| 1925 | ] |
| 1926 | assert [event.output_index for event in function_items] == [0, 1] |
| 1927 | |
| 1928 | argument_deltas = [ |
| 1929 | event.delta |
| 1930 | for event in events |
| 1931 | if event.type == "response.function_call_arguments.delta" |
| 1932 | ] |
| 1933 | assert argument_deltas == [ |
| 1934 | '{"location":"Vienna"}', |
| 1935 | '{"location":"Berlin"}', |
| 1936 | ] |
| 1937 | |
| 1938 | argument_done = [ |
| 1939 | event |
| 1940 | for event in events |
| 1941 | if event.type == "response.function_call_arguments.done" |
| 1942 | ] |
| 1943 | assert [event.arguments for event in argument_done] == [ |
| 1944 | '{"location":"Vienna"}', |
| 1945 | '{"location":"Berlin"}', |
| 1946 | ] |
| 1947 | assert [event.output_index for event in argument_done] == [0, 1] |
| 1948 | |
| 1949 | function_done = [ |
| 1950 | event |
| 1951 | for event in events |
| 1952 | if event.type == "response.output_item.done" |
| 1953 | and getattr(event.item, "type", None) == "function_call" |
| 1954 | ] |
| 1955 | assert [event.item.arguments for event in function_done] == [ |
| 1956 | '{"location":"Vienna"}', |
| 1957 | '{"location":"Berlin"}', |
| 1958 | ] |
| 1959 | assert [event.output_index for event in function_done] == [0, 1] |
| 1960 | |
| 1961 | @pytest.mark.skip_global_cleanup |
| 1962 | @pytest.mark.asyncio |
| 1963 | async def test_auto_tool_choice_first_delta_tool_call_does_not_duplicate_item( |
| 1964 | self, monkeypatch |
| 1965 | ): |
| 1966 | monkeypatch.setattr(envs, "VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", False) |
| 1967 | |
| 1968 | delta_sequence = [ |
| 1969 | DeltaMessage( |
| 1970 | tool_calls=[ |
| 1971 | DeltaToolCall( |
| 1972 | id="call_test", |
| 1973 | type="function", |
| 1974 | index=0, |
| 1975 | function=DeltaFunctionCall( |
| 1976 | name="get_weather", |
| 1977 | arguments="", |
| 1978 | ), |
| 1979 | ) |
| 1980 | ] |
| 1981 | ), |
| 1982 | DeltaMessage( |
| 1983 | tool_calls=[ |
| 1984 | DeltaToolCall( |
| 1985 | index=0, |
| 1986 | function=DeltaFunctionCall( |
| 1987 | arguments='{"location":"Berlin"}', |
| 1988 | ), |
| 1989 | ) |
| 1990 | ] |
| 1991 | ), |
| 1992 | ] |
| 1993 | events = await self._collect_events(delta_sequence) |
| 1994 | |
| 1995 | function_items = [ |
| 1996 | event |
| 1997 | for event in events |
| 1998 | if event.type == "response.output_item.added" |
| 1999 | and getattr(event.item, "type", None) == "function_call" |
| 2000 | ] |
| 2001 | assert len(function_items) == 1 |
| 2002 | assert function_items[0].item.name == "get_weather" |
| 2003 | |
| 2004 | argument_deltas = [ |
| 2005 | event.delta |
| 2006 | for event in events |
| 2007 | if event.type == "response.function_call_arguments.delta" |
| 2008 | ] |
| 2009 | assert "".join(argument_deltas) == '{"location":"Berlin"}' |
| 2010 | |
| 2011 | @pytest.mark.skip_global_cleanup |
| 2012 | @pytest.mark.asyncio |
| 2013 | async def test_compound_content_and_tool_name_args_same_delta(self, monkeypatch): |
| 2014 | monkeypatch.setattr(envs, "VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", False) |
| 2015 | |
| 2016 | tool_args = '{"location":"Berlin"}' |
| 2017 | |
| 2018 | delta_sequence = [ |
| 2019 | DeltaMessage( |
| 2020 | content="Let me check.", |
| 2021 | tool_calls=[ |
| 2022 | DeltaToolCall( |
| 2023 | id="call_weather", |
| 2024 | type="function", |
| 2025 | index=0, |
| 2026 | function=DeltaFunctionCall(name="get_weather"), |
| 2027 | ), |
| 2028 | DeltaToolCall( |
| 2029 | index=0, |
| 2030 | function=DeltaFunctionCall(arguments=tool_args), |
| 2031 | ), |
| 2032 | ], |
| 2033 | ) |
| 2034 | ] |
| 2035 | |
| 2036 | events = await self._collect_events(delta_sequence) |
| 2037 | |
| 2038 | text_deltas = [ |
| 2039 | event.delta |
| 2040 | for event in events |
| 2041 | if event.type == "response.output_text.delta" |
| 2042 | ] |
| 2043 | assert text_deltas == ["Let me check."] |
| 2044 | |
| 2045 | argument_deltas = [ |
| 2046 | event.delta |
| 2047 | for event in events |
| 2048 | if event.type == "response.function_call_arguments.delta" |
| 2049 | ] |
| 2050 | assert argument_deltas == [tool_args] |
| 2051 | |
| 2052 | types = [event.type for event in events] |
| 2053 | assert types.index("response.output_text.delta") < types.index( |
| 2054 | "response.function_call_arguments.delta" |
| 2055 | ) |
| 2056 | |
| 2057 | argument_done = [ |
| 2058 | event |
| 2059 | for event in events |
| 2060 | if event.type == "response.function_call_arguments.done" |
| 2061 | ] |
| 2062 | assert [event.arguments for event in argument_done] == [tool_args] |
| 2063 | |
| 2064 | function_done = [ |
| 2065 | event |
| 2066 | for event in events |
| 2067 | if event.type == "response.output_item.done" |
| 2068 | and getattr(event.item, "type", None) == "function_call" |
| 2069 | ] |
| 2070 | assert len(function_done) == 1 |
| 2071 | assert function_done[0].item.name == "get_weather" |
| 2072 | assert function_done[0].item.arguments == tool_args |
| 2073 | |
| 2074 | |
| 2075 | @pytest.mark.skip_global_cleanup |
| 2076 | @pytest.mark.asyncio |
| 2077 | async def test_stream_completed_response_reuses_streamed_items(monkeypatch): |
| 2078 | """response.completed must carry the streamed items (same ids, |
| 2079 | call_id, logprobs) rather than a reparse of the full output.""" |
| 2080 | monkeypatch.setattr(envs, "VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", False) |
| 2081 | serving = _make_serving_instance_with_reasoning() |
| 2082 | call = DeltaToolCall( |
| 2083 | id="chatcmpl-tool-parser-id", |
| 2084 | index=0, |
| 2085 | function=DeltaFunctionCall(name="get_weather", arguments="{}"), |
| 2086 | ) |
| 2087 | response_parser = _mock_parser_with_reasoning( |
| 2088 | serving, |
| 2089 | [ |
| 2090 | DeltaMessage(reasoning="think"), |
| 2091 | DeltaMessage(content="Hi"), |
| 2092 | DeltaMessage(tool_calls=[call]), |
| 2093 | ], |
| 2094 | ) |
| 2095 | response_parser.count_reasoning_tokens = MagicMock(return_value=1) |
| 2096 | context = SimpleContext(response_parser=response_parser) |
| 2097 | |
| 2098 | async def result_generator(): |
| 2099 | for token_id, text in [(10, "think"), (20, "Hi"), (30, "call")]: |
| 2100 | output = _make_request_output(text, [token_id]) |
| 2101 | output.outputs[0].logprobs = [ |
| 2102 | {token_id: SampleLogprob(logprob=-0.5, decoded_token=text)} |
| 2103 | ] |
| 2104 | context.append_output(output) |
| 2105 | yield context |
| 2106 | |
| 2107 | events = [ |
| 2108 | event |
| 2109 | async for event in serving.responses_stream_generator( |
| 2110 | request=ResponsesRequest( |
| 2111 | input="hi", |
| 2112 | include=["message.output_text.logprobs"], |
| 2113 | stream=True, |
| 2114 | store=False, |
| 2115 | ), |
| 2116 | sampling_params=SamplingParams(max_tokens=16), |
| 2117 | result_generator=result_generator(), |
| 2118 | context=context, |
| 2119 | model_name="test-model", |
| 2120 | tokenizer=MagicMock(), |
| 2121 | request_metadata=RequestResponseMetadata(request_id="req"), |
| 2122 | ) |
| 2123 | ] |
| 2124 | |
| 2125 | streamed = [e.item for e in events if e.type == "response.output_item.done"] |
| 2126 | assert events[-1].response.output == streamed |
| 2127 | _, message, function_call = streamed |
| 2128 | assert [lp.token for lp in message.content[0].logprobs] == ["Hi"] |
| 2129 | assert function_call.call_id == "chatcmpl-tool-parser-id" |
| 2130 | |
| 2131 | |
| 2132 | @pytest.mark.asyncio |
| 2133 | @pytest.mark.parametrize(("parallel_tool_calls", "num_calls"), [(False, 1), (None, 3)]) |
| 2134 | async def test_streaming_parallel_tool_calls( |
| 2135 | monkeypatch, parallel_tool_calls, num_calls |
| 2136 | ): |
| 2137 | """parallel_tool_calls=False keeps only the first call even when the grammar |
| 2138 | does not limit the model; None keeps every call.""" |
| 2139 | monkeypatch.setattr(envs, "VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", False) |
| 2140 | serving = _make_serving_instance(tool_parser="hermes") |
| 2141 | request = ResponsesRequest( |
| 2142 | input="hi", |
| 2143 | tools=[{"type": "function", "name": "get_weather", "parameters": {}}], |
| 2144 | parallel_tool_calls=parallel_tool_calls, |
| 2145 | stream=True, |
| 2146 | ) |
| 2147 | context = SimpleContext( |
| 2148 | response_parser=serving._make_response_parser(request, MagicMock(), {}) |
| 2149 | ) |
| 2150 | args = [f'{{"city": "{city}"}}' for city in ("Paris", "Berlin", "Tokyo")] |
| 2151 | |
| 2152 | async def result_generator(): |
| 2153 | for i, arg in enumerate(args): |
| 2154 | text = f'<tool_call>\n{{"name": "get_weather", "arguments": {arg}}}' |
| 2155 | context.append_output(_make_request_output(f"{text}\n</tool_call>\n", [i])) |
| 2156 | yield context |
| 2157 | |
| 2158 | events = [ |
| 2159 | event |
| 2160 | async for event in serving.responses_stream_generator( |
| 2161 | request=request, |
| 2162 | sampling_params=SamplingParams(max_tokens=64), |
| 2163 | result_generator=result_generator(), |
| 2164 | context=context, |
| 2165 | model_name="test-model", |
| 2166 | tokenizer=MagicMock(), |
| 2167 | request_metadata=RequestResponseMetadata(request_id="req"), |
| 2168 | ) |
| 2169 | ] |
| 2170 | |
| 2171 | assert [ |
| 2172 | event.output_index |
| 2173 | for event in events |
| 2174 | if event.type == "response.output_item.added" |
| 2175 | ] == list(range(num_calls)) |
| 2176 | assert all(getattr(event, "output_index", 0) < num_calls for event in events) |
| 2177 | assert [ |
| 2178 | event.item.arguments |
| 2179 | for event in events |
| 2180 | if event.type == "response.output_item.done" |
| 2181 | ] == args[:num_calls] |
| 2182 | assert [item.arguments for item in events[-1].response.output] == args[:num_calls] |