| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from dataclasses import dataclass, field |
| 5 | from typing import Any |
| 6 | from unittest.mock import AsyncMock, MagicMock |
| 7 | |
| 8 | import pytest |
| 9 | |
| 10 | from vllm.config.multimodal import MultiModalConfig |
| 11 | from vllm.entrypoints.generate.base.protocol import RequestResponseMetadata |
| 12 | from vllm.entrypoints.openai.completion.protocol import CompletionRequest |
| 13 | from vllm.entrypoints.openai.completion.serving import OpenAIServingCompletion |
| 14 | from vllm.entrypoints.openai.models.protocol import BaseModelPath |
| 15 | from vllm.entrypoints.openai.models.serving import OpenAIServingModels |
| 16 | from vllm.entrypoints.scale_out.render.serving import ServingRender |
| 17 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 18 | from vllm.exceptions import GenerationError, VLLMValidationError |
| 19 | from vllm.outputs import CompletionOutput, RequestOutput |
| 20 | from vllm.renderers.deepseek_v4 import DeepseekV4Renderer |
| 21 | from vllm.renderers.hf import HfRenderer |
| 22 | from vllm.renderers.online_renderer import OnlineRenderer |
| 23 | from vllm.tokenizers.registry import cached_tokenizer_from_config |
| 24 | from vllm.v1.engine.async_llm import AsyncLLM |
| 25 | from vllm.v1.metrics.stats import RequestSpecDecodeMetrics, RequestStateStats |
| 26 | |
| 27 | MODEL_NAME = "openai-community/gpt2" |
| 28 | MODEL_NAME_SHORT = "gpt2" |
| 29 | DEEPSEEK_V4_FIM_BEGIN = "<｜fim▁begin｜>" |
| 30 | DEEPSEEK_V4_FIM_HOLE = "<｜fim▁hole｜>" |
| 31 | DEEPSEEK_V4_FIM_END = "<｜fim▁end｜>" |
| 32 | _PER_REQUEST_STATS = RequestStateStats( |
| 33 | queued_ts=1.0, |
| 34 | scheduled_ts=1.5, |
| 35 | first_token_ts=2.0, |
| 36 | last_token_ts=3.0, |
| 37 | num_generation_tokens=2, |
| 38 | ) |
| 39 | BASE_MODEL_PATHS = [ |
| 40 | BaseModelPath(name=MODEL_NAME, model_path=MODEL_NAME), |
| 41 | BaseModelPath(name=MODEL_NAME_SHORT, model_path=MODEL_NAME_SHORT), |
| 42 | ] |
| 43 | |
| 44 | |
| 45 | @dataclass |
| 46 | class MockHFConfig: |
| 47 | model_type: str = "any" |
| 48 | |
| 49 | |
| 50 | @dataclass |
| 51 | class MockModelConfig: |
| 52 | task = "generate" |
| 53 | runner_type = "generate" |
| 54 | model = MODEL_NAME |
| 55 | tokenizer = MODEL_NAME |
| 56 | trust_remote_code = False |
| 57 | tokenizer_mode = "auto" |
| 58 | max_model_len = 100 |
| 59 | revision = None |
| 60 | code_revision = None |
| 61 | tokenizer_revision = None |
| 62 | multimodal_config = MultiModalConfig() |
| 63 | hf_config = MockHFConfig() |
| 64 | logits_processors: list[str] | None = None |
| 65 | diff_sampling_param: dict | None = None |
| 66 | allowed_local_media_path: str = "" |
| 67 | allowed_media_domains: list[str] | None = None |
| 68 | encoder_config = None |
| 69 | generation_config: str = "auto" |
| 70 | media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict) |
| 71 | skip_tokenizer_init: bool = False |
| 72 | is_encoder_decoder: bool = False |
| 73 | is_multimodal_model: bool = False |
| 74 | supports_multimodal_inputs: bool = False |
| 75 | renderer_num_workers: int = 1 |
| 76 | |
| 77 | def get_diff_sampling_param(self): |
| 78 | return self.diff_sampling_param or {} |
| 79 | |
| 80 | |
| 81 | @dataclass |
| 82 | class MockParallelConfig: |
| 83 | _api_process_rank: int = 0 |
| 84 | |
| 85 | |
| 86 | @dataclass |
| 87 | class MockVllmConfig: |
| 88 | model_config: MockModelConfig |
| 89 | parallel_config: MockParallelConfig |
| 90 | |
| 91 | |
| 92 | def _build_serving_completion(engine: AsyncLLM) -> OpenAIServingCompletion: |
| 93 | models = OpenAIServingModels( |
| 94 | engine_client=engine, |
| 95 | base_model_paths=BASE_MODEL_PATHS, |
| 96 | ) |
| 97 | online_renderer = OnlineRenderer( |
| 98 | model_config=engine.model_config, |
| 99 | renderer=engine.renderer, |
| 100 | request_logger=None, |
| 101 | chat_template=None, |
| 102 | chat_template_content_format="auto", |
| 103 | ) |
| 104 | return OpenAIServingCompletion( |
| 105 | engine, |
| 106 | models, |
| 107 | online_renderer=online_renderer, |
| 108 | request_logger=None, |
| 109 | ) |
| 110 | |
| 111 | |
| 112 | def _build_minimal_metrics_serving_completion( |
| 113 | enable_per_request_metrics: bool, |
| 114 | ) -> OpenAIServingCompletion: |
| 115 | serving = OpenAIServingCompletion.__new__(OpenAIServingCompletion) |
| 116 | serving.enable_prompt_tokens_details = False |
| 117 | serving.system_fingerprint = None |
| 118 | serving.enable_per_request_metrics = enable_per_request_metrics |
| 119 | return serving |
| 120 | |
| 121 | |
| 122 | def _make_metrics_request_output( |
| 123 | metrics: RequestStateStats | None = _PER_REQUEST_STATS, |
| 124 | ) -> RequestOutput: |
| 125 | return RequestOutput( |
| 126 | request_id="test-id", |
| 127 | prompt="Test prompt", |
| 128 | prompt_token_ids=[1, 2, 3], |
| 129 | prompt_logprobs=None, |
| 130 | outputs=[ |
| 131 | CompletionOutput( |
| 132 | index=0, |
| 133 | text="Hello", |
| 134 | token_ids=[100, 101], |
| 135 | cumulative_logprob=None, |
| 136 | logprobs=None, |
| 137 | finish_reason="stop", |
| 138 | ) |
| 139 | ], |
| 140 | finished=True, |
| 141 | metrics=metrics, |
| 142 | ) |
| 143 | |
| 144 | |
| 145 | def _build_renderer(model_config: MockModelConfig): |
| 146 | return HfRenderer( |
| 147 | MockVllmConfig(model_config, parallel_config=MockParallelConfig()), |
| 148 | cached_tokenizer_from_config(model_config), |
| 149 | ) |
| 150 | |
| 151 | |
| 152 | def test_completion_per_request_metrics_follow_server_flag(): |
| 153 | request = CompletionRequest(model=MODEL_NAME, prompt="Test prompt", max_tokens=10) |
| 154 | request_output = _make_metrics_request_output() |
| 155 | |
| 156 | disabled_serving = _build_minimal_metrics_serving_completion( |
| 157 | enable_per_request_metrics=False |
| 158 | ) |
| 159 | disabled_response = disabled_serving.request_output_to_completion_response( |
| 160 | [request_output], |
| 161 | request, |
| 162 | "cmpl-test-id", |
| 163 | 0, |
| 164 | MODEL_NAME, |
| 165 | None, |
| 166 | RequestResponseMetadata(request_id="cmpl-test-id"), |
| 167 | ) |
| 168 | assert disabled_response.metrics is None |
| 169 | |
| 170 | enabled_serving = _build_minimal_metrics_serving_completion( |
| 171 | enable_per_request_metrics=True |
| 172 | ) |
| 173 | enabled_response = enabled_serving.request_output_to_completion_response( |
| 174 | [request_output], |
| 175 | request, |
| 176 | "cmpl-test-id", |
| 177 | 0, |
| 178 | MODEL_NAME, |
| 179 | None, |
| 180 | RequestResponseMetadata(request_id="cmpl-test-id"), |
| 181 | ) |
| 182 | assert enabled_response.metrics is not None |
| 183 | assert enabled_response.metrics.time_to_first_token_ms == pytest.approx(500.0) |
| 184 | |
| 185 | |
| 186 | def test_completion_per_request_metrics_suppressed_for_multiple_prompts(): |
| 187 | serving = _build_minimal_metrics_serving_completion(enable_per_request_metrics=True) |
| 188 | response = serving.request_output_to_completion_response( |
| 189 | [_make_metrics_request_output(), _make_metrics_request_output()], |
| 190 | CompletionRequest( |
| 191 | model=MODEL_NAME, |
| 192 | prompt=["Test prompt", "Another prompt"], |
| 193 | max_tokens=10, |
| 194 | ), |
| 195 | "cmpl-test-id", |
| 196 | 0, |
| 197 | MODEL_NAME, |
| 198 | None, |
| 199 | RequestResponseMetadata(request_id="cmpl-test-id"), |
| 200 | ) |
| 201 | assert response.metrics is None |
| 202 | |
| 203 | |
| 204 | def _spec_decode_metrics() -> RequestSpecDecodeMetrics: |
| 205 | # Two verify steps: accept 3 drafts, then 1 -> histogram [0, 1, 0, 1]. |
| 206 | m = RequestSpecDecodeMetrics.new(num_spec_tokens=3) |
| 207 | m.observe(num_draft_tokens=3, num_accepted=3) |
| 208 | m.observe(num_draft_tokens=3, num_accepted=1) |
| 209 | return m |
| 210 | |
| 211 | |
| 212 | def _make_spec_decode_request_output( |
| 213 | num_seqs: int = 1, with_metrics: bool = True |
| 214 | ) -> RequestOutput: |
| 215 | outputs = [ |
| 216 | CompletionOutput( |
| 217 | index=i, |
| 218 | text="Hello", |
| 219 | token_ids=[100, 101], |
| 220 | cumulative_logprob=None, |
| 221 | logprobs=None, |
| 222 | finish_reason="stop", |
| 223 | spec_decode_metrics=_spec_decode_metrics() if with_metrics else None, |
| 224 | ) |
| 225 | for i in range(num_seqs) |
| 226 | ] |
| 227 | return RequestOutput( |
| 228 | request_id="test-id", |
| 229 | prompt="Test prompt", |
| 230 | prompt_token_ids=[1, 2, 3], |
| 231 | prompt_logprobs=None, |
| 232 | outputs=outputs, |
| 233 | finished=True, |
| 234 | metrics=None, |
| 235 | ) |
| 236 | |
| 237 | |
| 238 | def _completion_response(serving, request, request_output): |
| 239 | return serving.request_output_to_completion_response( |
| 240 | [request_output], |
| 241 | request, |
| 242 | "cmpl-test-id", |
| 243 | 0, |
| 244 | MODEL_NAME, |
| 245 | None, |
| 246 | RequestResponseMetadata(request_id="cmpl-test-id"), |
| 247 | ) |
| 248 | |
| 249 | |
| 250 | def test_completion_spec_decode_metrics_present_for_single_sequence(): |
| 251 | # Timing off, but the sequence carries acceptance metrics -> the metrics |
| 252 | # object is created just to hold metrics.speculative_decoding. |
| 253 | serving = _build_minimal_metrics_serving_completion( |
| 254 | enable_per_request_metrics=False |
| 255 | ) |
| 256 | response = _completion_response( |
| 257 | serving, |
| 258 | CompletionRequest(model=MODEL_NAME, prompt="Test prompt", max_tokens=10), |
| 259 | _make_spec_decode_request_output(num_seqs=1), |
| 260 | ) |
| 261 | assert response.metrics is not None |
| 262 | assert response.metrics.time_to_first_token_ms is None # timing not requested |
| 263 | spec = response.metrics.speculative_decoding |
| 264 | assert spec is not None |
| 265 | assert spec.acceptance_histogram == [0, 1, 0, 1] # dense, index j |
| 266 | assert spec.num_spec_steps == 2 |
| 267 | assert spec.num_spec_tokens == 3 |
| 268 | assert spec.mean_acceptance_length == pytest.approx(3.0) # 1 + (3 + 1) / 2 |
| 269 | |
| 270 | |
| 271 | def test_completion_spec_decode_metrics_suppressed_for_n_gt_1(): |
| 272 | # Per-request metrics can't be attributed to one of the n sequences. |
| 273 | serving = _build_minimal_metrics_serving_completion( |
| 274 | enable_per_request_metrics=False |
| 275 | ) |
| 276 | response = _completion_response( |
| 277 | serving, |
| 278 | CompletionRequest(model=MODEL_NAME, prompt="Test prompt", n=2, max_tokens=10), |
| 279 | _make_spec_decode_request_output(num_seqs=2), |
| 280 | ) |
| 281 | assert response.metrics is None |
| 282 | |
| 283 | |
| 284 | def test_completion_spec_decode_metrics_absent_when_not_collected(): |
| 285 | # Flag off -> the sequence carries no acceptance metrics -> no metrics object. |
| 286 | serving = _build_minimal_metrics_serving_completion( |
| 287 | enable_per_request_metrics=False |
| 288 | ) |
| 289 | response = _completion_response( |
| 290 | serving, |
| 291 | CompletionRequest(model=MODEL_NAME, prompt="Test prompt", max_tokens=10), |
| 292 | _make_spec_decode_request_output(num_seqs=1, with_metrics=False), |
| 293 | ) |
| 294 | assert response.metrics is None |
| 295 | |
| 296 | |
| 297 | def test_completion_metrics_carries_both_timing_and_spec_decode(): |
| 298 | serving = _build_minimal_metrics_serving_completion(enable_per_request_metrics=True) |
| 299 | request_output = _make_spec_decode_request_output(num_seqs=1) |
| 300 | request_output.metrics = _PER_REQUEST_STATS # timing source |
| 301 | response = _completion_response( |
| 302 | serving, |
| 303 | CompletionRequest(model=MODEL_NAME, prompt="Test prompt", max_tokens=10), |
| 304 | request_output, |
| 305 | ) |
| 306 | assert response.metrics is not None |
| 307 | assert response.metrics.time_to_first_token_ms == pytest.approx(500.0) |
| 308 | assert response.metrics.speculative_decoding is not None |
| 309 | assert response.metrics.speculative_decoding.num_spec_steps == 2 |
| 310 | |
| 311 | |
| 312 | def _build_online_renderer_for_completion( |
| 313 | model_config: MockModelConfig, |
| 314 | renderer: Any | None = None, |
| 315 | ) -> OnlineRenderer: |
| 316 | if renderer is None: |
| 317 | renderer = MagicMock() |
| 318 | renderer.render_completion_suffix.return_value = None |
| 319 | |
| 320 | return OnlineRenderer( |
| 321 | model_config=model_config, |
| 322 | renderer=renderer, |
| 323 | request_logger=None, |
| 324 | chat_template=None, |
| 325 | chat_template_content_format="auto", |
| 326 | ) |
| 327 | |
| 328 | |
| 329 | def _deepseek_v4_suffix_renderer( |
| 330 | model_config: MockModelConfig, |
| 331 | ) -> DeepseekV4Renderer: |
| 332 | return DeepseekV4Renderer( |
| 333 | MockVllmConfig(model_config, parallel_config=MockParallelConfig()), |
| 334 | None, |
| 335 | ) |
| 336 | |
| 337 | |
| 338 | @pytest.mark.asyncio |
| 339 | @pytest.mark.skip_global_cleanup |
| 340 | async def test_completion_suffix_uses_renderer_capability(): |
| 341 | model_config = MockModelConfig() |
| 342 | online_renderer = _build_online_renderer_for_completion( |
| 343 | model_config, |
| 344 | renderer=_deepseek_v4_suffix_renderer(model_config), |
| 345 | ) |
| 346 | online_renderer.preprocess_completion = AsyncMock(return_value=[{"ok": True}]) |
| 347 | |
| 348 | request = CompletionRequest( |
| 349 | model=MODEL_NAME, |
| 350 | prompt="def fib(n):\n return ", |
| 351 | suffix="\n\nprint(fib(10))", |
| 352 | max_tokens=64, |
| 353 | ) |
| 354 | |
| 355 | result = await online_renderer.render_completion(request) |
| 356 | |
| 357 | assert result == [{"ok": True}] |
| 358 | online_renderer.preprocess_completion.assert_awaited_once() |
| 359 | assert online_renderer.preprocess_completion.call_args.kwargs["prompt_input"] == ( |
| 360 | f"{DEEPSEEK_V4_FIM_BEGIN}def fib(n):\n return " |
| 361 | f"{DEEPSEEK_V4_FIM_HOLE}\n\nprint(fib(10))" |
| 362 | f"{DEEPSEEK_V4_FIM_END}" |
| 363 | ) |
| 364 | |
| 365 | |
| 366 | @pytest.mark.asyncio |
| 367 | @pytest.mark.skip_global_cleanup |
| 368 | async def test_completion_suffix_supports_text_prompt_list(): |
| 369 | model_config = MockModelConfig() |
| 370 | online_renderer = _build_online_renderer_for_completion( |
| 371 | model_config, |
| 372 | renderer=_deepseek_v4_suffix_renderer(model_config), |
| 373 | ) |
| 374 | online_renderer.preprocess_completion = AsyncMock(return_value=[{"ok": True}]) |
| 375 | |
| 376 | request = CompletionRequest( |
| 377 | model=MODEL_NAME, |
| 378 | prompt=["prefix A", "prefix B"], |
| 379 | suffix=" suffix", |
| 380 | ) |
| 381 | |
| 382 | result = await online_renderer.render_completion(request) |
| 383 | |
| 384 | assert result == [{"ok": True}] |
| 385 | assert online_renderer.preprocess_completion.call_args.kwargs["prompt_input"] == [ |
| 386 | f"{DEEPSEEK_V4_FIM_BEGIN}prefix A" |
| 387 | f"{DEEPSEEK_V4_FIM_HOLE} suffix{DEEPSEEK_V4_FIM_END}", |
| 388 | f"{DEEPSEEK_V4_FIM_BEGIN}prefix B" |
| 389 | f"{DEEPSEEK_V4_FIM_HOLE} suffix{DEEPSEEK_V4_FIM_END}", |
| 390 | ] |
| 391 | |
| 392 | |
| 393 | @pytest.mark.asyncio |
| 394 | @pytest.mark.skip_global_cleanup |
| 395 | async def test_completion_suffix_rejects_renderer_without_fim_support(): |
| 396 | model_config = MockModelConfig() |
| 397 | online_renderer = _build_online_renderer_for_completion(model_config) |
| 398 | online_renderer.preprocess_completion = AsyncMock() |
| 399 | |
| 400 | request = CompletionRequest( |
| 401 | model=MODEL_NAME, |
| 402 | prompt="prefix", |
| 403 | suffix="suffix", |
| 404 | ) |
| 405 | |
| 406 | result = await online_renderer.render_completion(request) |
| 407 | |
| 408 | assert isinstance(result, ErrorResponse) |
| 409 | assert "FIM completion rendering" in result.error.message |
| 410 | online_renderer.preprocess_completion.assert_not_awaited() |
| 411 | |
| 412 | |
| 413 | @pytest.mark.asyncio |
| 414 | @pytest.mark.skip_global_cleanup |
| 415 | async def test_completion_suffix_rejects_echo(): |
| 416 | model_config = MockModelConfig() |
| 417 | online_renderer = _build_online_renderer_for_completion( |
| 418 | model_config, |
| 419 | renderer=_deepseek_v4_suffix_renderer(model_config), |
| 420 | ) |
| 421 | online_renderer.preprocess_completion = AsyncMock() |
| 422 | |
| 423 | request = CompletionRequest( |
| 424 | model=MODEL_NAME, |
| 425 | prompt="prefix", |
| 426 | suffix="suffix", |
| 427 | echo=True, |
| 428 | ) |
| 429 | |
| 430 | result = await online_renderer.render_completion(request) |
| 431 | |
| 432 | assert isinstance(result, ErrorResponse) |
| 433 | assert "Echo is unsupported with suffix" in result.error.message |
| 434 | online_renderer.preprocess_completion.assert_not_awaited() |
| 435 | |
| 436 | |
| 437 | @pytest.mark.asyncio |
| 438 | @pytest.mark.skip_global_cleanup |
| 439 | async def test_completion_suffix_rejects_prompt_embeds(): |
| 440 | model_config = MockModelConfig() |
| 441 | online_renderer = _build_online_renderer_for_completion( |
| 442 | model_config, |
| 443 | renderer=_deepseek_v4_suffix_renderer(model_config), |
| 444 | ) |
| 445 | online_renderer.preprocess_completion = AsyncMock() |
| 446 | |
| 447 | request = CompletionRequest( |
| 448 | model=MODEL_NAME, |
| 449 | prompt="prefix", |
| 450 | suffix="suffix", |
| 451 | prompt_embeds=b"embeds", |
| 452 | ) |
| 453 | |
| 454 | result = await online_renderer.render_completion(request) |
| 455 | |
| 456 | assert isinstance(result, ErrorResponse) |
| 457 | assert "suffix is not supported with prompt_embeds" in result.error.message |
| 458 | online_renderer.preprocess_completion.assert_not_awaited() |
| 459 | |
| 460 | |
| 461 | @pytest.mark.asyncio |
| 462 | @pytest.mark.skip_global_cleanup |
| 463 | @pytest.mark.parametrize("prompt", ([1, 2, 3], [[1, 2, 3]])) |
| 464 | async def test_completion_suffix_rejects_token_prompt(prompt): |
| 465 | model_config = MockModelConfig() |
| 466 | online_renderer = _build_online_renderer_for_completion( |
| 467 | model_config, |
| 468 | renderer=_deepseek_v4_suffix_renderer(model_config), |
| 469 | ) |
| 470 | online_renderer.preprocess_completion = AsyncMock() |
| 471 | |
| 472 | request = CompletionRequest( |
| 473 | model=MODEL_NAME, |
| 474 | prompt=prompt, |
| 475 | suffix="suffix", |
| 476 | ) |
| 477 | |
| 478 | result = await online_renderer.render_completion(request) |
| 479 | |
| 480 | assert isinstance(result, ErrorResponse) |
| 481 | assert "requires text prompt input" in result.error.message |
| 482 | online_renderer.preprocess_completion.assert_not_awaited() |
| 483 | |
| 484 | |
| 485 | @pytest.mark.asyncio |
| 486 | async def test_completion_error_non_stream(): |
| 487 | """Test finish_reason='error' returns 500 InternalServerError (non-streaming).""" |
| 488 | mock_engine = MagicMock(spec=AsyncLLM) |
| 489 | mock_engine.errored = False |
| 490 | mock_engine.model_config = MockModelConfig() |
| 491 | mock_engine.input_processor = MagicMock() |
| 492 | mock_engine.renderer = _build_renderer(mock_engine.model_config) |
| 493 | |
| 494 | serving_completion = _build_serving_completion(mock_engine) |
| 495 | |
| 496 | completion_output = CompletionOutput( |
| 497 | index=0, |
| 498 | text="", |
| 499 | token_ids=[], |
| 500 | cumulative_logprob=None, |
| 501 | logprobs=None, |
| 502 | finish_reason="error", |
| 503 | ) |
| 504 | |
| 505 | request_output = RequestOutput( |
| 506 | request_id="test-id", |
| 507 | prompt="Test prompt", |
| 508 | prompt_token_ids=[1, 2, 3], |
| 509 | prompt_logprobs=None, |
| 510 | outputs=[completion_output], |
| 511 | finished=True, |
| 512 | metrics=None, |
| 513 | lora_request=None, |
| 514 | encoder_prompt=None, |
| 515 | encoder_prompt_token_ids=None, |
| 516 | ) |
| 517 | |
| 518 | async def mock_generate(*args, **kwargs): |
| 519 | yield request_output |
| 520 | |
| 521 | mock_engine.generate = MagicMock(side_effect=mock_generate) |
| 522 | |
| 523 | request = CompletionRequest( |
| 524 | model=MODEL_NAME, |
| 525 | prompt="Test prompt", |
| 526 | max_tokens=10, |
| 527 | stream=False, |
| 528 | ) |
| 529 | |
| 530 | with pytest.raises(GenerationError): |
| 531 | await serving_completion.create_completion(request) |
| 532 | |
| 533 | |
| 534 | @pytest.mark.asyncio |
| 535 | async def test_openai_completion_keeps_mm_cache_for_engine_execution(): |
| 536 | mock_engine = MagicMock(spec=AsyncLLM) |
| 537 | mock_engine.errored = False |
| 538 | mock_engine.model_config = MockModelConfig() |
| 539 | mock_engine.input_processor = MagicMock() |
| 540 | mock_engine.renderer = _build_renderer(mock_engine.model_config) |
| 541 | |
| 542 | serving_completion = _build_serving_completion(mock_engine) |
| 543 | serving_completion.online_renderer.preprocess_completion = AsyncMock( |
| 544 | return_value=[{"prompt_token_ids": [1, 2, 3]}] |
| 545 | ) |
| 546 | |
| 547 | request = CompletionRequest( |
| 548 | model=MODEL_NAME, |
| 549 | prompt="Test prompt", |
| 550 | ) |
| 551 | |
| 552 | result = await serving_completion.render_completion_request(request) |
| 553 | |
| 554 | assert isinstance(result, list) |
| 555 | assert ( |
| 556 | serving_completion.online_renderer.preprocess_completion.call_args.kwargs[ |
| 557 | "skip_mm_cache" |
| 558 | ] |
| 559 | is False |
| 560 | ) |
| 561 | |
| 562 | |
| 563 | def _build_serving_render(engine: AsyncLLM) -> ServingRender: |
| 564 | models = OpenAIServingModels( |
| 565 | engine_client=engine, |
| 566 | base_model_paths=BASE_MODEL_PATHS, |
| 567 | ) |
| 568 | online_renderer = OnlineRenderer( |
| 569 | model_config=engine.model_config, |
| 570 | renderer=engine.renderer, |
| 571 | request_logger=None, |
| 572 | chat_template=None, |
| 573 | chat_template_content_format="auto", |
| 574 | ) |
| 575 | |
| 576 | serving_render = ServingRender(models, online_renderer) |
| 577 | |
| 578 | async def _fake_preprocess_chat(*args, **kwargs): |
| 579 | # return conversation, engine_inputs |
| 580 | return ( |
| 581 | [{"role": "user", "content": "Test"}], |
| 582 | [{"prompt_token_ids": [1, 2, 3]}], |
| 583 | ) |
| 584 | |
| 585 | serving_render.online_renderer.preprocess_chat = AsyncMock( |
| 586 | side_effect=_fake_preprocess_chat |
| 587 | ) |
| 588 | return serving_render |
| 589 | |
| 590 | |
| 591 | @pytest.mark.asyncio |
| 592 | async def test_renderer_only_completion_request_skips_mm_cache(): |
| 593 | mock_engine = MagicMock(spec=AsyncLLM) |
| 594 | mock_engine.errored = False |
| 595 | mock_engine.model_config = MockModelConfig() |
| 596 | mock_engine.input_processor = MagicMock() |
| 597 | mock_engine.renderer = _build_renderer(mock_engine.model_config) |
| 598 | |
| 599 | serving_render = _build_serving_render(mock_engine) |
| 600 | |
| 601 | serving_render.online_renderer.preprocess_completion = AsyncMock( |
| 602 | return_value=[{"prompt_token_ids": [1, 2, 3]}] |
| 603 | ) |
| 604 | |
| 605 | request = CompletionRequest( |
| 606 | model=MODEL_NAME, |
| 607 | prompt="Test prompt", |
| 608 | ) |
| 609 | |
| 610 | result = await serving_render.render_completion_request(request) |
| 611 | |
| 612 | assert isinstance(result, list) |
| 613 | assert ( |
| 614 | serving_render.online_renderer.preprocess_completion.call_args.kwargs[ |
| 615 | "skip_mm_cache" |
| 616 | ] |
| 617 | is True |
| 618 | ) |
| 619 | |
| 620 | |
| 621 | @pytest.mark.asyncio |
| 622 | async def test_completion_error_stream(): |
| 623 | """Test finish_reason='error' returns 500 InternalServerError (streaming).""" |
| 624 | mock_engine = MagicMock(spec=AsyncLLM) |
| 625 | mock_engine.errored = False |
| 626 | mock_engine.model_config = MockModelConfig() |
| 627 | mock_engine.input_processor = MagicMock() |
| 628 | mock_engine.renderer = _build_renderer(mock_engine.model_config) |
| 629 | |
| 630 | serving_completion = _build_serving_completion(mock_engine) |
| 631 | |
| 632 | completion_output_1 = CompletionOutput( |
| 633 | index=0, |
| 634 | text="Hello", |
| 635 | token_ids=[100], |
| 636 | cumulative_logprob=None, |
| 637 | logprobs=None, |
| 638 | finish_reason=None, |
| 639 | ) |
| 640 | |
| 641 | request_output_1 = RequestOutput( |
| 642 | request_id="test-id", |
| 643 | prompt="Test prompt", |
| 644 | prompt_token_ids=[1, 2, 3], |
| 645 | prompt_logprobs=None, |
| 646 | outputs=[completion_output_1], |
| 647 | finished=False, |
| 648 | metrics=None, |
| 649 | lora_request=None, |
| 650 | encoder_prompt=None, |
| 651 | encoder_prompt_token_ids=None, |
| 652 | ) |
| 653 | |
| 654 | completion_output_2 = CompletionOutput( |
| 655 | index=0, |
| 656 | text="Hello", |
| 657 | token_ids=[100], |
| 658 | cumulative_logprob=None, |
| 659 | logprobs=None, |
| 660 | finish_reason="error", |
| 661 | ) |
| 662 | |
| 663 | request_output_2 = RequestOutput( |
| 664 | request_id="test-id", |
| 665 | prompt="Test prompt", |
| 666 | prompt_token_ids=[1, 2, 3], |
| 667 | prompt_logprobs=None, |
| 668 | outputs=[completion_output_2], |
| 669 | finished=True, |
| 670 | metrics=None, |
| 671 | lora_request=None, |
| 672 | encoder_prompt=None, |
| 673 | encoder_prompt_token_ids=None, |
| 674 | ) |
| 675 | |
| 676 | async def mock_generate(*args, **kwargs): |
| 677 | yield request_output_1 |
| 678 | yield request_output_2 |
| 679 | |
| 680 | mock_engine.generate = MagicMock(side_effect=mock_generate) |
| 681 | |
| 682 | request = CompletionRequest( |
| 683 | model=MODEL_NAME, |
| 684 | prompt="Test prompt", |
| 685 | max_tokens=10, |
| 686 | stream=True, |
| 687 | ) |
| 688 | |
| 689 | response = await serving_completion.create_completion(request) |
| 690 | |
| 691 | chunks = [] |
| 692 | async for chunk in response: |
| 693 | chunks.append(chunk) |
| 694 | |
| 695 | assert len(chunks) >= 2 |
| 696 | assert any("Internal server error" in chunk for chunk in chunks), ( |
| 697 | f"Expected error message in chunks: {chunks}" |
| 698 | ) |
| 699 | assert chunks[-1] == "data: [DONE]\n\n" |
| 700 | |
| 701 | |
| 702 | def test_json_schema_response_format_missing_schema(): |
| 703 | """When response_format type is 'json_schema' but the json_schema field |
| 704 | is not provided, request construction should raise a validation error |
| 705 | so the API returns 400 instead of 500.""" |
| 706 | with pytest.raises(Exception, match="json_schema.*must be provided"): |
| 707 | CompletionRequest( |
| 708 | model=MODEL_NAME, |
| 709 | prompt="Test prompt", |
| 710 | max_tokens=10, |
| 711 | response_format={"type": "json_schema"}, |
| 712 | ) |
| 713 | |
| 714 | |
| 715 | @pytest.mark.parametrize("format_value", [None, {}]) |
| 716 | def test_structural_tag_response_format_invalid(format_value): |
| 717 | """Malformed structural tags should be rejected during request validation.""" |
| 718 | with pytest.raises( |
| 719 | VLLMValidationError, |
| 720 | match="Invalid response_format structural_tag", |
| 721 | ): |
| 722 | CompletionRequest( |
| 723 | model=MODEL_NAME, |
| 724 | prompt="Test prompt", |
| 725 | max_tokens=10, |
| 726 | response_format={"type": "structural_tag", "format": format_value}, |
| 727 | ) |
| 728 | |
| 729 | |
| 730 | @pytest.mark.parametrize("structural_tag", ["not json", ""]) |
| 731 | def test_structured_outputs_structural_tag_invalid(structural_tag): |
| 732 | """Malformed direct structured_outputs structural tags should be rejected.""" |
| 733 | with pytest.raises( |
| 734 | VLLMValidationError, |
| 735 | match="Invalid structured_outputs structural_tag", |
| 736 | ): |
| 737 | CompletionRequest( |
| 738 | model=MODEL_NAME, |
| 739 | prompt="Test prompt", |
| 740 | max_tokens=10, |
| 741 | structured_outputs={"structural_tag": structural_tag}, |
| 742 | ) |
| 743 | |
| 744 | |
| 745 | def test_negative_prompt_token_ids_nested(): |
| 746 | """Negative token IDs in prompt (nested list) should raise validation error.""" |
| 747 | with pytest.raises(Exception, match="greater than or equal to 0"): |
| 748 | CompletionRequest( |
| 749 | model=MODEL_NAME, |
| 750 | prompt=[[-1]], |
| 751 | max_tokens=10, |
| 752 | ) |
| 753 | |
| 754 | |
| 755 | def test_negative_prompt_token_ids_flat(): |
| 756 | """Negative token IDs in prompt (flat list) should raise validation error.""" |
| 757 | with pytest.raises(Exception, match="greater than or equal to 0"): |
| 758 | CompletionRequest( |
| 759 | model=MODEL_NAME, |
| 760 | prompt=[-1], |
| 761 | max_tokens=10, |
| 762 | ) |
| 763 | |
| 764 | |
| 765 | def test_logprobs_minus_one_allowed(): |
| 766 | """logprobs=-1 means "return all logprobs". The sampling layer and the chat |
| 767 | top_logprobs / prompt_logprobs validators all accept -1, so the completion |
| 768 | logprobs validator must accept it too instead of rejecting it as negative.""" |
| 769 | request = CompletionRequest( |
| 770 | model=MODEL_NAME, |
| 771 | prompt="Test prompt", |
| 772 | max_tokens=10, |
| 773 | logprobs=-1, |
| 774 | ) |
| 775 | assert request.logprobs == -1 |
| 776 | |
| 777 | |
| 778 | def test_logprobs_below_minus_one_rejected(): |
| 779 | """Values more negative than -1 stay invalid.""" |
| 780 | with pytest.raises(Exception, match="must be a positive value or -1"): |
| 781 | CompletionRequest( |
| 782 | model=MODEL_NAME, |
| 783 | prompt="Test prompt", |
| 784 | max_tokens=10, |
| 785 | logprobs=-2, |
| 786 | ) |
| 787 | |
| 788 | |
| 789 | class TestCompletionPromptListLimit: |
| 790 | """Regression tests for CVE: unbounded prompt list fan-out.""" |
| 791 | |
| 792 | def test_scalar_prompt_allowed(self): |
| 793 | request = CompletionRequest( |
| 794 | model=MODEL_NAME, |
| 795 | prompt="hello", |
| 796 | max_tokens=1, |
| 797 | ) |
| 798 | assert request.prompt == "hello" |
| 799 | |
| 800 | def test_single_token_list_allowed(self): |
| 801 | request = CompletionRequest( |
| 802 | model=MODEL_NAME, |
| 803 | prompt=[1, 2, 3], |
| 804 | max_tokens=1, |
| 805 | ) |
| 806 | assert request.prompt == [1, 2, 3] |
| 807 | |
| 808 | def test_bounded_text_prompt_list_allowed(self, monkeypatch): |
| 809 | monkeypatch.setenv("VLLM_MAX_COMPLETION_PROMPTS", "10") |
| 810 | from vllm import envs |
| 811 | |
| 812 | if hasattr(envs.__getattr__, "cache_clear"): |
| 813 | envs.__getattr__.cache_clear() |
| 814 | |
| 815 | request = CompletionRequest( |
| 816 | model=MODEL_NAME, |
| 817 | prompt=["a", "b", "c"], |
| 818 | max_tokens=1, |
| 819 | ) |
| 820 | assert request.prompt == ["a", "b", "c"] |
| 821 | |
| 822 | def test_bounded_token_id_prompt_list_allowed(self, monkeypatch): |
| 823 | monkeypatch.setenv("VLLM_MAX_COMPLETION_PROMPTS", "10") |
| 824 | from vllm import envs |
| 825 | |
| 826 | if hasattr(envs.__getattr__, "cache_clear"): |
| 827 | envs.__getattr__.cache_clear() |
| 828 | |
| 829 | request = CompletionRequest( |
| 830 | model=MODEL_NAME, |
| 831 | prompt=[[1], [2], [3]], |
| 832 | max_tokens=1, |
| 833 | ) |
| 834 | assert request.prompt == [[1], [2], [3]] |
| 835 | |
| 836 | def test_oversized_text_prompt_list_rejected(self, monkeypatch): |
| 837 | monkeypatch.setenv("VLLM_MAX_COMPLETION_PROMPTS", "5") |
| 838 | from vllm import envs |
| 839 | |
| 840 | if hasattr(envs.__getattr__, "cache_clear"): |
| 841 | envs.__getattr__.cache_clear() |
| 842 | |
| 843 | with pytest.raises( |
| 844 | Exception, match="prompt list length 10 exceeds the maximum" |
| 845 | ): |
| 846 | CompletionRequest( |
| 847 | model=MODEL_NAME, |
| 848 | prompt=["x"] * 10, |
| 849 | max_tokens=1, |
| 850 | ) |
| 851 | |
| 852 | def test_oversized_token_id_prompt_list_rejected(self, monkeypatch): |
| 853 | monkeypatch.setenv("VLLM_MAX_COMPLETION_PROMPTS", "5") |
| 854 | from vllm import envs |
| 855 | |
| 856 | if hasattr(envs.__getattr__, "cache_clear"): |
| 857 | envs.__getattr__.cache_clear() |
| 858 | |
| 859 | with pytest.raises( |
| 860 | Exception, match="prompt list length 10 exceeds the maximum" |
| 861 | ): |
| 862 | CompletionRequest( |
| 863 | model=MODEL_NAME, |
| 864 | prompt=[[1]] * 10, |
| 865 | max_tokens=1, |
| 866 | ) |
| 867 | |
| 868 | def test_exact_limit_allowed(self, monkeypatch): |
| 869 | monkeypatch.setenv("VLLM_MAX_COMPLETION_PROMPTS", "5") |
| 870 | from vllm import envs |
| 871 | |
| 872 | if hasattr(envs.__getattr__, "cache_clear"): |
| 873 | envs.__getattr__.cache_clear() |
| 874 | |
| 875 | request = CompletionRequest( |
| 876 | model=MODEL_NAME, |
| 877 | prompt=["x"] * 5, |
| 878 | max_tokens=1, |
| 879 | ) |
| 880 | assert len(request.prompt) == 5 |
| 881 | |
| 882 | def test_one_over_limit_rejected(self, monkeypatch): |
| 883 | monkeypatch.setenv("VLLM_MAX_COMPLETION_PROMPTS", "5") |
| 884 | from vllm import envs |
| 885 | |
| 886 | if hasattr(envs.__getattr__, "cache_clear"): |
| 887 | envs.__getattr__.cache_clear() |
| 888 | |
| 889 | with pytest.raises(Exception, match="prompt list length 6 exceeds the maximum"): |
| 890 | CompletionRequest( |
| 891 | model=MODEL_NAME, |
| 892 | prompt=["x"] * 6, |
| 893 | max_tokens=1, |
| 894 | ) |
| 895 | |
| 896 | def test_oversized_prompt_embeds_list_rejected(self, monkeypatch): |
| 897 | monkeypatch.setenv("VLLM_MAX_COMPLETION_PROMPTS", "5") |
| 898 | from vllm import envs |
| 899 | |
| 900 | if hasattr(envs.__getattr__, "cache_clear"): |
| 901 | envs.__getattr__.cache_clear() |
| 902 | |
| 903 | with pytest.raises(Exception, match="prompt_embeds list length 10 exceeds"): |
| 904 | CompletionRequest( |
| 905 | model=MODEL_NAME, |
| 906 | prompt_embeds=[b"\x00"] * 10, |
| 907 | max_tokens=1, |
| 908 | ) |
| 909 | |
| 910 | def test_bounded_prompt_embeds_list_allowed(self, monkeypatch): |
| 911 | monkeypatch.setenv("VLLM_MAX_COMPLETION_PROMPTS", "5") |
| 912 | from vllm import envs |
| 913 | |
| 914 | if hasattr(envs.__getattr__, "cache_clear"): |
| 915 | envs.__getattr__.cache_clear() |
| 916 | |
| 917 | request = CompletionRequest( |
| 918 | model=MODEL_NAME, |
| 919 | prompt_embeds=[b"\x00"] * 5, |
| 920 | max_tokens=1, |
| 921 | ) |
| 922 | assert len(request.prompt_embeds) == 5 |
| 923 | |
| 924 | |
| 925 | @pytest.mark.parametrize("field_name", ["prompt_logprobs", "logprobs"]) |
| 926 | def test_non_numeric_logprobs_rejected(field_name): |
| 927 | """A non-numeric logprobs value must be a clean 400 validation error, not a |
| 928 | TypeError from the mode='before' comparison (which surfaces as HTTP 500).""" |
| 929 | with pytest.raises(VLLMValidationError, match=f"`{field_name}` must be an integer"): |
| 930 | CompletionRequest( |
| 931 | model=MODEL_NAME, |
| 932 | prompt="Test prompt", |
| 933 | max_tokens=10, |
| 934 | **{field_name: "2"}, |
| 935 | ) |