| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from types import SimpleNamespace |
| 5 | from unittest.mock import Mock |
| 6 | |
| 7 | import pytest |
| 8 | |
| 9 | from vllm import LLM, SamplingParams |
| 10 | from vllm.entrypoints.chat_utils import ChatCompletionMessageParam |
| 11 | from vllm.inputs import TextPrompt |
| 12 | from vllm.outputs import RequestOutput |
| 13 | |
| 14 | |
| 15 | def _make_mock_llm() -> LLM: |
| 16 | llm = object.__new__(LLM) |
| 17 | llm.model_config = SimpleNamespace( |
| 18 | runner_type="generate", enable_prompt_embeds=False |
| 19 | ) |
| 20 | return llm |
| 21 | |
| 22 | |
| 23 | def test_generate_forwards_mm_processor_kwargs(monkeypatch: pytest.MonkeyPatch) -> None: |
| 24 | llm = _make_mock_llm() |
| 25 | mm_processor_kwargs = {"num_crops": 4} |
| 26 | sampling_params = SamplingParams(max_tokens=1) |
| 27 | |
| 28 | _run_completion = Mock(return_value=["ok"]) |
| 29 | monkeypatch.setattr(llm, "_run_completion", _run_completion) |
| 30 | |
| 31 | outputs = llm.generate( |
| 32 | "prompt", |
| 33 | sampling_params=sampling_params, |
| 34 | mm_processor_kwargs=mm_processor_kwargs, |
| 35 | ) |
| 36 | |
| 37 | assert outputs == ["ok"] |
| 38 | assert _run_completion.call_args.kwargs["mm_processor_kwargs"] == ( |
| 39 | mm_processor_kwargs |
| 40 | ) |
| 41 | |
| 42 | |
| 43 | def test_enqueue_forwards_mm_processor_kwargs(monkeypatch: pytest.MonkeyPatch) -> None: |
| 44 | llm = _make_mock_llm() |
| 45 | mm_processor_kwargs = {"do_resize": False} |
| 46 | sampling_params = SamplingParams(max_tokens=1) |
| 47 | |
| 48 | _add_completion_requests = Mock(return_value=["req-0"]) |
| 49 | monkeypatch.setattr(llm, "_add_completion_requests", _add_completion_requests) |
| 50 | |
| 51 | request_ids = llm.enqueue( |
| 52 | "prompt", |
| 53 | sampling_params=sampling_params, |
| 54 | use_tqdm=False, |
| 55 | mm_processor_kwargs=mm_processor_kwargs, |
| 56 | ) |
| 57 | |
| 58 | assert request_ids == ["req-0"] |
| 59 | assert _add_completion_requests.call_args.kwargs["mm_processor_kwargs"] == ( |
| 60 | mm_processor_kwargs |
| 61 | ) |
| 62 | |
| 63 | |
| 64 | def test_chat_forwards_mm_processor_kwargs(monkeypatch: pytest.MonkeyPatch) -> None: |
| 65 | llm = _make_mock_llm() |
| 66 | mm_processor_kwargs = {"do_pan_and_scan": True} |
| 67 | sampling_params = SamplingParams(max_tokens=1) |
| 68 | messages: list[ChatCompletionMessageParam] = [{"role": "user", "content": "hello"}] |
| 69 | |
| 70 | _run_chat = Mock(return_value=["ok"]) |
| 71 | monkeypatch.setattr(llm, "_run_chat", _run_chat) |
| 72 | |
| 73 | outputs = llm.chat( |
| 74 | messages, |
| 75 | sampling_params=sampling_params, |
| 76 | mm_processor_kwargs=mm_processor_kwargs, |
| 77 | ) |
| 78 | |
| 79 | assert outputs == ["ok"] |
| 80 | assert _run_chat.call_args.kwargs["mm_processor_kwargs"] == mm_processor_kwargs |
| 81 | |
| 82 | |
| 83 | def test_enqueue_chat_forwards_mm_processor_kwargs( |
| 84 | monkeypatch: pytest.MonkeyPatch, |
| 85 | ) -> None: |
| 86 | llm = _make_mock_llm() |
| 87 | mm_processor_kwargs = {"do_pan_and_scan": True} |
| 88 | sampling_params = SamplingParams(max_tokens=1) |
| 89 | messages: list[ChatCompletionMessageParam] = [{"role": "user", "content": "hello"}] |
| 90 | |
| 91 | _add_chat_requests = Mock(return_value=["req-0"]) |
| 92 | monkeypatch.setattr(llm, "_add_chat_requests", _add_chat_requests) |
| 93 | |
| 94 | request_ids = llm.enqueue_chat( |
| 95 | messages, |
| 96 | sampling_params=sampling_params, |
| 97 | use_tqdm=False, |
| 98 | mm_processor_kwargs=mm_processor_kwargs, |
| 99 | ) |
| 100 | |
| 101 | assert request_ids == ["req-0"] |
| 102 | assert _add_chat_requests.call_args.kwargs["mm_processor_kwargs"] == ( |
| 103 | mm_processor_kwargs |
| 104 | ) |
| 105 | |
| 106 | |
| 107 | def test_run_chat_forwards_mm_processor_kwargs(monkeypatch: pytest.MonkeyPatch) -> None: |
| 108 | llm = _make_mock_llm() |
| 109 | mm_processor_kwargs = {"num_crops": 8} |
| 110 | sampling_params = SamplingParams(max_tokens=1) |
| 111 | messages: list[ChatCompletionMessageParam] = [{"role": "user", "content": "hello"}] |
| 112 | sentinel_output = ["done"] |
| 113 | |
| 114 | _add_chat_requests = Mock() |
| 115 | monkeypatch.setattr(llm, "_add_chat_requests", _add_chat_requests) |
| 116 | _run_engine = Mock(return_value=sentinel_output) |
| 117 | monkeypatch.setattr(llm, "_run_engine", _run_engine) |
| 118 | |
| 119 | outputs = llm._run_chat( |
| 120 | messages=messages, |
| 121 | params=sampling_params, |
| 122 | output_type=RequestOutput, |
| 123 | use_tqdm=False, |
| 124 | mm_processor_kwargs=mm_processor_kwargs, |
| 125 | ) |
| 126 | |
| 127 | assert outputs == sentinel_output |
| 128 | assert _add_chat_requests.call_args.kwargs["mm_processor_kwargs"] == ( |
| 129 | mm_processor_kwargs |
| 130 | ) |
| 131 | |
| 132 | |
| 133 | def test_run_completion_forwards_mm_processor_kwargs( |
| 134 | monkeypatch: pytest.MonkeyPatch, |
| 135 | ) -> None: |
| 136 | llm = _make_mock_llm() |
| 137 | mm_processor_kwargs = {"min_pixels": 4 * 28 * 28} |
| 138 | sampling_params = SamplingParams(max_tokens=1) |
| 139 | sentinel_output = ["done"] |
| 140 | |
| 141 | _add_completion_requests = Mock() |
| 142 | monkeypatch.setattr(llm, "_add_completion_requests", _add_completion_requests) |
| 143 | _run_engine = Mock(return_value=sentinel_output) |
| 144 | monkeypatch.setattr(llm, "_run_engine", _run_engine) |
| 145 | |
| 146 | outputs = llm._run_completion( |
| 147 | prompts=["prompt"], |
| 148 | params=sampling_params, |
| 149 | output_type=RequestOutput, |
| 150 | use_tqdm=False, |
| 151 | mm_processor_kwargs=mm_processor_kwargs, |
| 152 | ) |
| 153 | |
| 154 | assert outputs == sentinel_output |
| 155 | assert _add_completion_requests.call_args.kwargs["mm_processor_kwargs"] == ( |
| 156 | mm_processor_kwargs |
| 157 | ) |
| 158 | |
| 159 | |
| 160 | def test_add_completion_requests_forwards_mm_processor_kwargs( |
| 161 | monkeypatch: pytest.MonkeyPatch, |
| 162 | ) -> None: |
| 163 | llm = _make_mock_llm() |
| 164 | mm_processor_kwargs = {"max_dynamic_patch": 4} |
| 165 | sampling_params = SamplingParams(max_tokens=1) |
| 166 | |
| 167 | _params_to_seq = Mock(return_value=[sampling_params]) |
| 168 | monkeypatch.setattr(llm, "_params_to_seq", _params_to_seq) |
| 169 | _lora_request_to_seq = Mock(return_value=[None]) |
| 170 | monkeypatch.setattr(llm, "_lora_request_to_seq", _lora_request_to_seq) |
| 171 | _priority_to_seq = Mock(return_value=[0]) |
| 172 | monkeypatch.setattr(llm, "_priority_to_seq", _priority_to_seq) |
| 173 | _preprocess_cmpl_one = Mock(return_value={"prompt_token_ids": [1]}) |
| 174 | monkeypatch.setattr(llm, "_preprocess_cmpl_one", _preprocess_cmpl_one) |
| 175 | |
| 176 | captured_prompts = [] |
| 177 | |
| 178 | def fake_render_and_add_requests(*, prompts, **_kwargs): |
| 179 | captured_prompts.extend(prompts) |
| 180 | return ["req-0"] |
| 181 | |
| 182 | _render_and_add_requests = Mock(side_effect=fake_render_and_add_requests) |
| 183 | monkeypatch.setattr(llm, "_render_and_add_requests", _render_and_add_requests) |
| 184 | |
| 185 | request_ids = llm._add_completion_requests( |
| 186 | prompts=["prompt"], |
| 187 | params=sampling_params, |
| 188 | use_tqdm=False, |
| 189 | mm_processor_kwargs=mm_processor_kwargs, |
| 190 | ) |
| 191 | |
| 192 | assert request_ids == ["req-0"] |
| 193 | _preprocess_cmpl_one.assert_called_once_with( |
| 194 | "prompt", |
| 195 | None, |
| 196 | mm_processor_kwargs=mm_processor_kwargs, |
| 197 | ) |
| 198 | assert captured_prompts == [{"prompt_token_ids": [1]}] |
| 199 | |
| 200 | |
| 201 | def test_preprocess_cmpl_applies_mm_processor_kwargs_to_renderer( |
| 202 | monkeypatch: pytest.MonkeyPatch, |
| 203 | ) -> None: |
| 204 | llm = _make_mock_llm() |
| 205 | mm_processor_kwargs = {"num_crops": 8} |
| 206 | prompt: TextPrompt = {"prompt": "<image>", "multi_modal_data": {"image": object()}} |
| 207 | |
| 208 | renderer = Mock() |
| 209 | renderer.default_cmpl_tok_params = Mock() |
| 210 | renderer.default_cmpl_tok_params.with_kwargs.return_value = "tok-params" |
| 211 | renderer.render_cmpl.return_value = ["engine-input"] |
| 212 | llm.renderer = renderer |
| 213 | |
| 214 | monkeypatch.setattr( |
| 215 | "vllm.entrypoints.offline_utils.parse_model_prompt", |
| 216 | lambda _model_config, parsed_prompt: parsed_prompt, |
| 217 | ) |
| 218 | |
| 219 | outputs = llm._preprocess_cmpl( |
| 220 | [prompt], |
| 221 | mm_processor_kwargs=mm_processor_kwargs, |
| 222 | ) |
| 223 | |
| 224 | assert outputs == ["engine-input"] |
| 225 | renderer.render_cmpl.assert_called_once_with( |
| 226 | [prompt], |
| 227 | "tok-params", |
| 228 | prompt_extras={"mm_processor_kwargs": mm_processor_kwargs}, |
| 229 | ) |
| 230 | |
| 231 | |
| 232 | def test_preprocess_cmpl_keeps_prompt_mm_processor_kwargs_when_no_override( |
| 233 | monkeypatch: pytest.MonkeyPatch, |
| 234 | ) -> None: |
| 235 | llm = _make_mock_llm() |
| 236 | prompt: TextPrompt = { |
| 237 | "prompt": "<image>", |
| 238 | "multi_modal_data": {"image": object()}, |
| 239 | "mm_processor_kwargs": {"num_crops": 2}, |
| 240 | } |
| 241 | |
| 242 | renderer = Mock() |
| 243 | renderer.default_cmpl_tok_params = Mock() |
| 244 | renderer.default_cmpl_tok_params.with_kwargs.return_value = "tok-params" |
| 245 | renderer.render_cmpl.return_value = ["engine-input"] |
| 246 | llm.renderer = renderer |
| 247 | |
| 248 | monkeypatch.setattr( |
| 249 | "vllm.entrypoints.offline_utils.parse_model_prompt", |
| 250 | lambda _model_config, parsed_prompt: parsed_prompt, |
| 251 | ) |
| 252 | |
| 253 | outputs = llm._preprocess_cmpl([prompt]) |
| 254 | |
| 255 | assert outputs == ["engine-input"] |
| 256 | renderer.render_cmpl.assert_called_once_with( |
| 257 | [prompt], |
| 258 | "tok-params", |
| 259 | prompt_extras=None, |
| 260 | ) |
| 261 | |
| 262 | |
| 263 | def test_preprocess_chat_applies_mm_processor_kwargs_to_renderer() -> None: |
| 264 | llm = _make_mock_llm() |
| 265 | mm_processor_kwargs = {"num_crops": 8} |
| 266 | messages: list[list[ChatCompletionMessageParam]] = [ |
| 267 | [{"role": "user", "content": "Describe this image."}] |
| 268 | ] |
| 269 | |
| 270 | renderer = Mock() |
| 271 | renderer.tokenizer = object() |
| 272 | renderer.default_chat_tok_params = Mock() |
| 273 | renderer.default_chat_tok_params.with_kwargs.return_value = "tok-params" |
| 274 | renderer.render_chat.return_value = (messages, ["engine-input"]) |
| 275 | llm.renderer = renderer |
| 276 | |
| 277 | outputs = llm._preprocess_chat( |
| 278 | messages, |
| 279 | mm_processor_kwargs=mm_processor_kwargs, |
| 280 | ) |
| 281 | |
| 282 | assert outputs == ["engine-input"] |
| 283 | call_args = renderer.render_chat.call_args |
| 284 | assert call_args.args[0] == messages |
| 285 | assert call_args.args[1].mm_processor_kwargs == mm_processor_kwargs |
| 286 | assert call_args.args[2] == "tok-params" |
| 287 | assert call_args.kwargs["prompt_extras"] == { |
| 288 | "mm_processor_kwargs": mm_processor_kwargs |
| 289 | } |
| 290 | |
| 291 | |
| 292 | def test_preprocess_chat_omits_mm_processor_kwargs_when_no_override() -> None: |
| 293 | llm = _make_mock_llm() |
| 294 | messages: list[list[ChatCompletionMessageParam]] = [ |
| 295 | [{"role": "user", "content": "Describe this image."}] |
| 296 | ] |
| 297 | |
| 298 | renderer = Mock() |
| 299 | renderer.tokenizer = object() |
| 300 | renderer.default_chat_tok_params = Mock() |
| 301 | renderer.default_chat_tok_params.with_kwargs.return_value = "tok-params" |
| 302 | renderer.render_chat.return_value = (messages, ["engine-input"]) |
| 303 | llm.renderer = renderer |
| 304 | |
| 305 | outputs = llm._preprocess_chat(messages) |
| 306 | |
| 307 | assert outputs == ["engine-input"] |
| 308 | call_args = renderer.render_chat.call_args |
| 309 | assert call_args.args[0] == messages |
| 310 | assert call_args.args[1].mm_processor_kwargs is None |
| 311 | assert call_args.args[2] == "tok-params" |
| 312 | assert call_args.kwargs["prompt_extras"] is None |
| 313 | |
| 314 | |
| 315 | def test_preprocess_chat_defaults_add_special_tokens_to_false() -> None: |
| 316 | # Matches `ChatCompletionRequest.add_special_tokens` on the server. |
| 317 | llm = _make_mock_llm() |
| 318 | messages: list[list[ChatCompletionMessageParam]] = [ |
| 319 | [{"role": "user", "content": "hi"}] |
| 320 | ] |
| 321 | |
| 322 | renderer = Mock() |
| 323 | renderer.tokenizer = object() |
| 324 | renderer.default_chat_tok_params = Mock() |
| 325 | renderer.default_chat_tok_params.with_kwargs.return_value = "tok-params" |
| 326 | renderer.render_chat.return_value = (messages, ["engine-input"]) |
| 327 | llm.renderer = renderer |
| 328 | |
| 329 | llm._preprocess_chat(messages) |
| 330 | |
| 331 | renderer.default_chat_tok_params.with_kwargs.assert_called_once_with( |
| 332 | add_special_tokens=False |
| 333 | ) |
| 334 | |
| 335 | |
| 336 | def test_preprocess_chat_tokenization_kwargs_override_add_special_tokens() -> None: |
| 337 | llm = _make_mock_llm() |
| 338 | messages: list[list[ChatCompletionMessageParam]] = [ |
| 339 | [{"role": "user", "content": "hi"}] |
| 340 | ] |
| 341 | |
| 342 | renderer = Mock() |
| 343 | renderer.tokenizer = object() |
| 344 | renderer.default_chat_tok_params = Mock() |
| 345 | renderer.default_chat_tok_params.with_kwargs.return_value = "tok-params" |
| 346 | renderer.render_chat.return_value = (messages, ["engine-input"]) |
| 347 | llm.renderer = renderer |
| 348 | |
| 349 | llm._preprocess_chat( |
| 350 | messages, |
| 351 | tokenization_kwargs={"add_special_tokens": True, "truncate_prompt_tokens": 8}, |
| 352 | ) |
| 353 | |
| 354 | renderer.default_chat_tok_params.with_kwargs.assert_called_once_with( |
| 355 | add_special_tokens=True, truncate_prompt_tokens=8 |
| 356 | ) |
| 357 | |
| 358 | |
| 359 | @pytest.fixture(scope="module") |
| 360 | def llava_llm(): |
| 361 | from vllm.config import ModelConfig, VllmConfig |
| 362 | from vllm.renderers.hf import HfRenderer |
| 363 | from vllm.tokenizers import cached_tokenizer_from_config |
| 364 | |
| 365 | # A real multimodal model (Llama tokenizer, BOS=1) so that the real |
| 366 | # processor default (`add_special_tokens=True`) is in play. Only |
| 367 | # config/tokenizer/processor files are fetched, no weights. |
| 368 | model_config = ModelConfig(model="llava-hf/llava-1.5-7b-hf", max_model_len=128) |
| 369 | renderer = HfRenderer( |
| 370 | VllmConfig(model_config=model_config), |
| 371 | cached_tokenizer_from_config(model_config), |
| 372 | ) |
| 373 | assert renderer.default_chat_tok_params.add_special_tokens is True |
| 374 | |
| 375 | llm = _make_mock_llm() |
| 376 | # `_preprocess_cmpl` parses prompts against the real model config. |
| 377 | llm.model_config = model_config |
| 378 | llm.renderer = renderer |
| 379 | return llm |
| 380 | |
| 381 | |
| 382 | class TestChatAddSpecialTokensDefault: |
| 383 | """`LLM.chat()` renders the chat template to text and then tokenizes it. |
| 384 | The multimodal processor default is `add_special_tokens=True`, so a |
| 385 | template that emits `bos_token` used to get a second BOS from the |
| 386 | tokenizer. `_preprocess_chat` now defaults `add_special_tokens=False` |
| 387 | like the online chat API does (`ChatCompletionRequest`), unless the |
| 388 | caller overrides it via `tokenization_kwargs`. |
| 389 | """ |
| 390 | |
| 391 | # Mirrors the Gemma 3 chat template and the bundled deepseek_vl2 / |
| 392 | # deepseek_ocr templates, whose rendered output starts with the BOS token. |
| 393 | BOS_TEMPLATE = ( |
| 394 | "{{ bos_token }}{% for m in messages %}{{ m['content'] }}{% endfor %}" |
| 395 | ) |
| 396 | MESSAGES = [[{"role": "user", "content": "hi"}]] |
| 397 | |
| 398 | def _chat_token_ids(self, llm, **kwargs): |
| 399 | (engine_input,) = llm._preprocess_chat( |
| 400 | self.MESSAGES, |
| 401 | chat_template=self.BOS_TEMPLATE, |
| 402 | add_generation_prompt=False, |
| 403 | **kwargs, |
| 404 | ) |
| 405 | return engine_input["prompt_token_ids"] |
| 406 | |
| 407 | def test_chat_does_not_duplicate_template_bos(self, llava_llm): |
| 408 | bos = llava_llm.renderer.tokenizer.bos_token_id |
| 409 | prompt_token_ids = self._chat_token_ids(llava_llm) |
| 410 | |
| 411 | assert prompt_token_ids[0] == bos |
| 412 | assert prompt_token_ids.count(bos) == 1 |
| 413 | |
| 414 | def test_chat_matches_online_chat_api(self, llava_llm): |
| 415 | from vllm.renderers.params import ChatParams, TokenizeParams |
| 416 | |
| 417 | # `ChatCompletionRequest.add_special_tokens` defaults to `False`. |
| 418 | online_tok_params = TokenizeParams( |
| 419 | max_total_tokens=llava_llm.renderer.model_config.max_model_len, |
| 420 | add_special_tokens=False, |
| 421 | ) |
| 422 | chat_params = ChatParams( |
| 423 | chat_template=self.BOS_TEMPLATE, |
| 424 | chat_template_kwargs=dict(tokenize=False, add_generation_prompt=False), |
| 425 | ) |
| 426 | _, (online_input,) = llava_llm.renderer.render_chat( |
| 427 | self.MESSAGES, chat_params, online_tok_params |
| 428 | ) |
| 429 | |
| 430 | assert self._chat_token_ids(llava_llm) == online_input["prompt_token_ids"] |
| 431 | |
| 432 | def test_explicit_tokenization_kwargs_override_default(self, llava_llm): |
| 433 | bos = llava_llm.renderer.tokenizer.bos_token_id |
| 434 | prompt_token_ids = self._chat_token_ids( |
| 435 | llava_llm, tokenization_kwargs={"add_special_tokens": True} |
| 436 | ) |
| 437 | |
| 438 | assert prompt_token_ids.count(bos) == 2 |
| 439 | |
| 440 | def test_generate_keeps_processor_default(self, llava_llm): |
| 441 | # Raw prompts have no chat template to emit BOS, so `LLM.generate()` |
| 442 | # must keep letting the tokenizer add it. |
| 443 | bos = llava_llm.renderer.tokenizer.bos_token_id |
| 444 | (engine_input,) = llava_llm._preprocess_cmpl(["hi"]) |
| 445 | |
| 446 | assert engine_input["prompt_token_ids"][0] == bos |