| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import asyncio |
| 5 | import warnings |
| 6 | from collections.abc import Mapping |
| 7 | from typing import Literal |
| 8 | from unittest.mock import MagicMock |
| 9 | |
| 10 | import pytest |
| 11 | import torch |
| 12 | |
| 13 | from vllm.assets.audio import AudioAsset |
| 14 | from vllm.assets.image import ImageAsset |
| 15 | from vllm.assets.video import VideoAsset |
| 16 | from vllm.config import ModelConfig |
| 17 | from vllm.entrypoints.chat_utils import ( |
| 18 | MEDIA_CONNECTOR_REGISTRY, |
| 19 | AsyncMultiModalItemTracker, |
| 20 | ChatCompletionMessageParam, |
| 21 | ConversationMessage, |
| 22 | _load_embeds_dict, |
| 23 | _parse_metadata_array, |
| 24 | _postprocess_messages, |
| 25 | parse_chat_messages, |
| 26 | parse_chat_messages_async, |
| 27 | validate_chat_template, |
| 28 | ) |
| 29 | from vllm.exceptions import VLLMValidationError |
| 30 | from vllm.inputs import MultiModalDataDict, MultiModalUUIDDict |
| 31 | from vllm.multimodal.utils import ( |
| 32 | encode_audio_url, |
| 33 | encode_image_url, |
| 34 | encode_video_url, |
| 35 | ) |
| 36 | from vllm.utils.serial_utils import tensor2base64 |
| 37 | |
| 38 | KIMI_K2_5_MODEL_ID = "moonshotai/Kimi-K2.5" |
| 39 | PHI3V_MODEL_ID = "microsoft/Phi-3.5-vision-instruct" |
| 40 | QWEN2AUDIO_MODEL_ID = "Qwen/Qwen2-Audio-7B-Instruct" |
| 41 | QWEN25OMNI_MODEL_ID = "Qwen/Qwen2.5-Omni-7B" |
| 42 | MISTRAL_MODEL_ID = "mistralai/Mistral-Small-3.1-24B-Instruct-2503" |
| 43 | |
| 44 | |
| 45 | @pytest.mark.parametrize("values", [[1, 32, 48], [0.0, 0.12345678912345678]]) |
| 46 | def test_json_metadata_preserves_numeric_precision(values): |
| 47 | tensor = _parse_metadata_array("grid", values, {"grid"}) |
| 48 | assert tensor.tolist() == values |
| 49 | assert tensor.dtype == ( |
| 50 | torch.float64 if isinstance(values[0], float) else torch.int64 |
| 51 | ) |
| 52 | |
| 53 | |
| 54 | @pytest.mark.parametrize( |
| 55 | "key,values", |
| 56 | [ |
| 57 | ("image_embeds", [1, 2]), |
| 58 | ("grid", [True]), |
| 59 | ("grid", [float("nan")]), |
| 60 | ("grid", [float("inf")]), |
| 61 | ("grid", [[1, 2]]), |
| 62 | ("grid", [2**64]), |
| 63 | ], |
| 64 | ) |
| 65 | def test_json_arrays_are_only_valid_numeric_metadata(key, values): |
| 66 | with pytest.raises(VLLMValidationError): |
| 67 | _parse_metadata_array(key, values, {"grid"}) |
| 68 | |
| 69 | |
| 70 | @pytest.mark.asyncio |
| 71 | async def test_json_metadata_bypasses_tensor_deserialization(): |
| 72 | from unittest.mock import AsyncMock |
| 73 | |
| 74 | tensor = torch.ones(2, 3) |
| 75 | fetch = AsyncMock(return_value=tensor) |
| 76 | result = await _load_embeds_dict( |
| 77 | {"image_embeds": "legacy-base64", "image_grid_thw": [1, 2, 2]}, fetch |
| 78 | ) |
| 79 | fetch.assert_awaited_once_with("legacy-base64") |
| 80 | assert result["image_embeds"] is tensor |
| 81 | assert result["image_grid_thw"] == [1, 2, 2] |
| 82 | |
| 83 | |
| 84 | @pytest.fixture(scope="function") |
| 85 | def kimi_k2_5_model_config(): |
| 86 | return ModelConfig( |
| 87 | KIMI_K2_5_MODEL_ID, |
| 88 | runner="generate", |
| 89 | trust_remote_code=True, |
| 90 | limit_mm_per_prompt={ |
| 91 | "image": 2, |
| 92 | }, |
| 93 | ) |
| 94 | |
| 95 | |
| 96 | @pytest.fixture(scope="function") |
| 97 | def phi3v_model_config(): |
| 98 | return ModelConfig( |
| 99 | PHI3V_MODEL_ID, |
| 100 | runner="generate", |
| 101 | trust_remote_code=True, |
| 102 | limit_mm_per_prompt={ |
| 103 | "image": 2, |
| 104 | }, |
| 105 | ) |
| 106 | |
| 107 | |
| 108 | @pytest.fixture(scope="function") |
| 109 | def phi3v_model_config_mm_interleaved(): |
| 110 | return ModelConfig( |
| 111 | PHI3V_MODEL_ID, |
| 112 | runner="generate", |
| 113 | trust_remote_code=True, |
| 114 | interleave_mm_strings=True, |
| 115 | limit_mm_per_prompt={ |
| 116 | "image": 2, |
| 117 | }, |
| 118 | ) |
| 119 | |
| 120 | |
| 121 | @pytest.fixture(scope="function") |
| 122 | def phi3v_model_config_image_embeds(): |
| 123 | return ModelConfig( |
| 124 | PHI3V_MODEL_ID, |
| 125 | runner="generate", |
| 126 | trust_remote_code=True, |
| 127 | limit_mm_per_prompt={ |
| 128 | "image": 2, |
| 129 | }, |
| 130 | enable_mm_embeds=True, |
| 131 | ) |
| 132 | |
| 133 | |
| 134 | @pytest.fixture(scope="function") |
| 135 | def qwen25omni_model_config_image_embeds(): |
| 136 | return ModelConfig( |
| 137 | QWEN25OMNI_MODEL_ID, |
| 138 | runner="generate", |
| 139 | limit_mm_per_prompt={"image": 2}, |
| 140 | enable_mm_embeds=True, |
| 141 | ) |
| 142 | |
| 143 | |
| 144 | @pytest.fixture(scope="function") |
| 145 | def qwen25omni_model_config_video_embeds(): |
| 146 | return ModelConfig( |
| 147 | QWEN25OMNI_MODEL_ID, |
| 148 | runner="generate", |
| 149 | limit_mm_per_prompt={"video": 2}, |
| 150 | enable_mm_embeds=True, |
| 151 | ) |
| 152 | |
| 153 | |
| 154 | @pytest.fixture(scope="function") |
| 155 | def qwen2_audio_model_config(): |
| 156 | return ModelConfig( |
| 157 | QWEN2AUDIO_MODEL_ID, |
| 158 | runner="generate", |
| 159 | trust_remote_code=True, |
| 160 | limit_mm_per_prompt={ |
| 161 | "audio": 1, |
| 162 | }, |
| 163 | ) |
| 164 | |
| 165 | |
| 166 | @pytest.fixture(scope="function") |
| 167 | def audio_embeds_model_config(): |
| 168 | return ModelConfig( |
| 169 | QWEN2AUDIO_MODEL_ID, |
| 170 | runner="generate", |
| 171 | trust_remote_code=True, |
| 172 | limit_mm_per_prompt={ |
| 173 | "audio": 2, |
| 174 | }, |
| 175 | enable_mm_embeds=True, |
| 176 | ) |
| 177 | |
| 178 | |
| 179 | @pytest.fixture(scope="function") |
| 180 | def qwen25omni_model_config_mm_interleaved(): |
| 181 | return ModelConfig( |
| 182 | QWEN25OMNI_MODEL_ID, |
| 183 | runner="generate", |
| 184 | interleave_mm_strings=True, |
| 185 | limit_mm_per_prompt={ |
| 186 | "image": 2, |
| 187 | "audio": 1, |
| 188 | "video": 1, |
| 189 | }, |
| 190 | ) |
| 191 | |
| 192 | |
| 193 | @pytest.fixture(scope="function") |
| 194 | def mistral_model_config(): |
| 195 | return ModelConfig( |
| 196 | MISTRAL_MODEL_ID, |
| 197 | runner="generate", |
| 198 | limit_mm_per_prompt={ |
| 199 | "image": 2, |
| 200 | }, |
| 201 | ) |
| 202 | |
| 203 | |
| 204 | @pytest.fixture(scope="module") |
| 205 | def image_url(): |
| 206 | image = ImageAsset("cherry_blossom") |
| 207 | return encode_image_url(image.pil_image) |
| 208 | |
| 209 | |
| 210 | @pytest.fixture(scope="module") |
| 211 | def video_url(): |
| 212 | video = VideoAsset("baby_reading", 1) |
| 213 | return encode_video_url(video.np_ndarrays) |
| 214 | |
| 215 | |
| 216 | @pytest.fixture(scope="module") |
| 217 | def audio_url(): |
| 218 | audio = AudioAsset("mary_had_lamb") |
| 219 | return encode_audio_url(*audio.audio_and_sample_rate) |
| 220 | |
| 221 | |
| 222 | def _assert_mm_data_is_image_input( |
| 223 | mm_data: MultiModalDataDict | None, |
| 224 | image_count: int, |
| 225 | skipped_image_indices: list | None = None, |
| 226 | ) -> None: |
| 227 | assert mm_data is not None |
| 228 | assert set(mm_data.keys()) == {"image"} |
| 229 | |
| 230 | image_data = mm_data.get("image") |
| 231 | assert image_data is not None |
| 232 | |
| 233 | assert isinstance(image_data, list) and len(image_data) == image_count |
| 234 | if skipped_image_indices is not None: |
| 235 | for i in skipped_image_indices: |
| 236 | assert image_data[i] is None |
| 237 | |
| 238 | |
| 239 | def _assert_mm_data_is_vision_chunk_input( |
| 240 | mm_data: MultiModalDataDict | None, |
| 241 | vision_chunk_count: int, |
| 242 | ) -> None: |
| 243 | assert mm_data is not None |
| 244 | assert set(mm_data.keys()) == {"vision_chunk"} |
| 245 | |
| 246 | vision_chunk_data = mm_data.get("vision_chunk") |
| 247 | assert vision_chunk_data is not None |
| 248 | |
| 249 | assert ( |
| 250 | isinstance(vision_chunk_data, list) |
| 251 | and len(vision_chunk_data) == vision_chunk_count |
| 252 | ) |
| 253 | |
| 254 | |
| 255 | def _assert_mm_uuids( |
| 256 | mm_uuids: MultiModalUUIDDict | None, |
| 257 | media_count: int, |
| 258 | expected_uuids: list[str | None], |
| 259 | modality: str = "image", |
| 260 | ) -> None: |
| 261 | if len(expected_uuids) > 0: |
| 262 | assert mm_uuids is not None |
| 263 | assert modality in mm_uuids |
| 264 | |
| 265 | image_uuids = mm_uuids.get(modality) |
| 266 | assert image_uuids is not None |
| 267 | |
| 268 | assert isinstance(image_uuids, list) and len(image_uuids) == media_count |
| 269 | |
| 270 | assert image_uuids == expected_uuids |
| 271 | else: |
| 272 | assert mm_uuids is None |
| 273 | |
| 274 | |
| 275 | ModalityType = Literal["image", "video", "audio"] |
| 276 | MultiModalDataCounts = Mapping[ModalityType, int] |
| 277 | |
| 278 | |
| 279 | def _assert_mm_data_inputs( |
| 280 | mm_data: MultiModalDataDict | None, |
| 281 | data_count: MultiModalDataCounts, |
| 282 | skipped_media_indices: dict[str, list] | None = None, # modality -> list[int] |
| 283 | ) -> None: |
| 284 | assert mm_data is not None |
| 285 | assert set(data_count.keys()) == (set(mm_data.keys())) |
| 286 | |
| 287 | for modality, n in data_count.items(): |
| 288 | modality_data = mm_data.get(modality) |
| 289 | assert modality_data is not None |
| 290 | assert isinstance(modality_data, list) and len(modality_data) == n |
| 291 | |
| 292 | if skipped_media_indices is not None: |
| 293 | skipped_media_indices_for_modality = skipped_media_indices.get(modality) |
| 294 | assert skipped_media_indices_for_modality is not None |
| 295 | for i in skipped_media_indices_for_modality: |
| 296 | assert modality_data[i] is None |
| 297 | |
| 298 | |
| 299 | def test_parse_chat_messages_single_image( |
| 300 | phi3v_model_config, |
| 301 | image_url, |
| 302 | ): |
| 303 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 304 | [ |
| 305 | { |
| 306 | "role": "user", |
| 307 | "content": [ |
| 308 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 309 | {"type": "text", "text": "What's in the image?"}, |
| 310 | ], |
| 311 | } |
| 312 | ], |
| 313 | phi3v_model_config, |
| 314 | content_format="string", |
| 315 | ) |
| 316 | |
| 317 | assert conversation == [ |
| 318 | {"role": "user", "content": "<|image_1|>\nWhat's in the image?"} |
| 319 | ] |
| 320 | _assert_mm_data_is_image_input(mm_data, 1) |
| 321 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None]) |
| 322 | |
| 323 | |
| 324 | def test_parse_chat_messages_single_image_with_uuid( |
| 325 | phi3v_model_config, |
| 326 | image_url, |
| 327 | ): |
| 328 | image_uuid = str(hash(image_url)) |
| 329 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 330 | [ |
| 331 | { |
| 332 | "role": "user", |
| 333 | "content": [ |
| 334 | { |
| 335 | "type": "image_url", |
| 336 | "image_url": { |
| 337 | "url": image_url, |
| 338 | }, |
| 339 | "uuid": image_uuid, |
| 340 | }, |
| 341 | {"type": "text", "text": "What's in the image?"}, |
| 342 | ], |
| 343 | } |
| 344 | ], |
| 345 | phi3v_model_config, |
| 346 | content_format="string", |
| 347 | ) |
| 348 | |
| 349 | assert conversation == [ |
| 350 | {"role": "user", "content": "<|image_1|>\nWhat's in the image?"} |
| 351 | ] |
| 352 | _assert_mm_data_is_image_input(mm_data, 1) |
| 353 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[image_uuid]) |
| 354 | |
| 355 | |
| 356 | def test_parse_chat_messages_single_empty_image_with_uuid( |
| 357 | phi3v_model_config, |
| 358 | image_url, |
| 359 | ): |
| 360 | image_uuid = str(hash(image_url)) |
| 361 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 362 | [ |
| 363 | { |
| 364 | "role": "user", |
| 365 | "content": [ |
| 366 | { |
| 367 | "type": "image_url", |
| 368 | "image_url": None, |
| 369 | "uuid": image_uuid, |
| 370 | }, |
| 371 | {"type": "text", "text": "What's in the image?"}, |
| 372 | ], |
| 373 | } |
| 374 | ], |
| 375 | phi3v_model_config, |
| 376 | content_format="string", |
| 377 | ) |
| 378 | |
| 379 | assert conversation == [ |
| 380 | {"role": "user", "content": "<|image_1|>\nWhat's in the image?"} |
| 381 | ] |
| 382 | _assert_mm_data_is_image_input(mm_data, 1, skipped_image_indices=[0]) |
| 383 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[image_uuid]) |
| 384 | |
| 385 | |
| 386 | def test_parse_chat_messages_single_image_with_bad_uuid_format( |
| 387 | phi3v_model_config, |
| 388 | image_url, |
| 389 | ): |
| 390 | image_uuid = str(hash(image_url)) |
| 391 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 392 | [ |
| 393 | { |
| 394 | "role": "user", |
| 395 | "content": [ |
| 396 | { |
| 397 | "type": "image_url", |
| 398 | "image_url": { |
| 399 | "url": image_url, |
| 400 | "uuid": image_uuid, |
| 401 | }, |
| 402 | "bad_uuid_key": image_uuid, |
| 403 | }, |
| 404 | {"type": "text", "text": "What's in the image?"}, |
| 405 | ], |
| 406 | } |
| 407 | ], |
| 408 | phi3v_model_config, |
| 409 | content_format="string", |
| 410 | ) |
| 411 | |
| 412 | assert conversation == [ |
| 413 | {"role": "user", "content": "<|image_1|>\nWhat's in the image?"} |
| 414 | ] |
| 415 | _assert_mm_data_is_image_input(mm_data, 1) |
| 416 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None]) |
| 417 | |
| 418 | |
| 419 | def test_parse_chat_messages_multiple_images_with_uuids( |
| 420 | phi3v_model_config, |
| 421 | image_url, |
| 422 | ): |
| 423 | image_uuid1 = "my_uuid_1" |
| 424 | image_uuid2 = "my_uuid_2" |
| 425 | |
| 426 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 427 | [ |
| 428 | { |
| 429 | "role": "user", |
| 430 | "content": [ |
| 431 | { |
| 432 | "type": "image_url", |
| 433 | "image_url": { |
| 434 | "url": image_url, |
| 435 | }, |
| 436 | "uuid": image_uuid1, |
| 437 | }, |
| 438 | { |
| 439 | "type": "image_url", |
| 440 | "image_url": { |
| 441 | "url": image_url, |
| 442 | }, |
| 443 | "uuid": image_uuid2, |
| 444 | }, |
| 445 | {"type": "text", "text": "What's in the image?"}, |
| 446 | ], |
| 447 | } |
| 448 | ], |
| 449 | phi3v_model_config, |
| 450 | content_format="string", |
| 451 | ) |
| 452 | |
| 453 | assert conversation == [ |
| 454 | { |
| 455 | "role": "user", |
| 456 | "content": "<|image_1|>\n<|image_2|>\nWhat's in the image?", |
| 457 | } |
| 458 | ] |
| 459 | _assert_mm_data_is_image_input(mm_data, 2) |
| 460 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid1, image_uuid2]) |
| 461 | |
| 462 | |
| 463 | def test_parse_chat_messages_multiple_empty_images_with_uuids( |
| 464 | phi3v_model_config, |
| 465 | image_url, |
| 466 | ): |
| 467 | image_uuid1 = "my_uuid_1" |
| 468 | image_uuid2 = "my_uuid_2" |
| 469 | |
| 470 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 471 | [ |
| 472 | { |
| 473 | "role": "user", |
| 474 | "content": [ |
| 475 | { |
| 476 | "type": "image_url", |
| 477 | "image_url": None, |
| 478 | "uuid": image_uuid1, |
| 479 | }, |
| 480 | { |
| 481 | "type": "image_url", |
| 482 | "image_url": None, |
| 483 | "uuid": image_uuid2, |
| 484 | }, |
| 485 | {"type": "text", "text": "What's in the image?"}, |
| 486 | ], |
| 487 | } |
| 488 | ], |
| 489 | phi3v_model_config, |
| 490 | content_format="string", |
| 491 | ) |
| 492 | |
| 493 | assert conversation == [ |
| 494 | { |
| 495 | "role": "user", |
| 496 | "content": "<|image_1|>\n<|image_2|>\nWhat's in the image?", |
| 497 | } |
| 498 | ] |
| 499 | _assert_mm_data_is_image_input(mm_data, 2, skipped_image_indices=[0, 1]) |
| 500 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid1, image_uuid2]) |
| 501 | |
| 502 | |
| 503 | def test_parse_chat_messages_mixed_empty_images_with_uuids( |
| 504 | phi3v_model_config, |
| 505 | image_url, |
| 506 | ): |
| 507 | image_uuid1 = "my_uuid_1" |
| 508 | image_uuid2 = "my_uuid_2" |
| 509 | |
| 510 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 511 | [ |
| 512 | { |
| 513 | "role": "user", |
| 514 | "content": [ |
| 515 | { |
| 516 | "type": "image_url", |
| 517 | "image_url": { |
| 518 | "url": image_url, |
| 519 | }, |
| 520 | "uuid": image_uuid1, |
| 521 | }, |
| 522 | { |
| 523 | "type": "image_url", |
| 524 | "image_url": None, |
| 525 | "uuid": image_uuid2, |
| 526 | }, |
| 527 | {"type": "text", "text": "What's in the image?"}, |
| 528 | ], |
| 529 | } |
| 530 | ], |
| 531 | phi3v_model_config, |
| 532 | content_format="string", |
| 533 | ) |
| 534 | |
| 535 | assert conversation == [ |
| 536 | { |
| 537 | "role": "user", |
| 538 | "content": "<|image_1|>\n<|image_2|>\nWhat's in the image?", |
| 539 | } |
| 540 | ] |
| 541 | _assert_mm_data_is_image_input(mm_data, 2, skipped_image_indices=[1]) |
| 542 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid1, image_uuid2]) |
| 543 | |
| 544 | |
| 545 | @pytest.mark.asyncio |
| 546 | async def test_parse_chat_messages_single_image_with_uuid_async( |
| 547 | phi3v_model_config, |
| 548 | image_url, |
| 549 | ): |
| 550 | image_uuid = str(hash(image_url)) |
| 551 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 552 | [ |
| 553 | { |
| 554 | "role": "user", |
| 555 | "content": [ |
| 556 | { |
| 557 | "type": "image_url", |
| 558 | "image_url": {"url": image_url}, |
| 559 | "uuid": image_uuid, |
| 560 | }, |
| 561 | {"type": "text", "text": "What's in the image?"}, |
| 562 | ], |
| 563 | } |
| 564 | ], |
| 565 | phi3v_model_config, |
| 566 | content_format="string", |
| 567 | ) |
| 568 | |
| 569 | assert conversation == [ |
| 570 | {"role": "user", "content": "<|image_1|>\nWhat's in the image?"} |
| 571 | ] |
| 572 | _assert_mm_data_is_image_input(mm_data, 1) |
| 573 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[image_uuid]) |
| 574 | |
| 575 | |
| 576 | @pytest.mark.asyncio |
| 577 | async def test_parse_chat_messages_empty_image_with_uuid_async( |
| 578 | phi3v_model_config, |
| 579 | image_url, |
| 580 | ): |
| 581 | image_uuid = str(hash(image_url)) |
| 582 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 583 | [ |
| 584 | { |
| 585 | "role": "user", |
| 586 | "content": [ |
| 587 | { |
| 588 | "type": "image_url", |
| 589 | "image_url": None, |
| 590 | "uuid": image_uuid, |
| 591 | }, |
| 592 | {"type": "text", "text": "What's in the image?"}, |
| 593 | ], |
| 594 | } |
| 595 | ], |
| 596 | phi3v_model_config, |
| 597 | content_format="string", |
| 598 | ) |
| 599 | |
| 600 | assert conversation == [ |
| 601 | {"role": "user", "content": "<|image_1|>\nWhat's in the image?"} |
| 602 | ] |
| 603 | _assert_mm_data_is_image_input(mm_data, 1, skipped_image_indices=[0]) |
| 604 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[image_uuid]) |
| 605 | |
| 606 | |
| 607 | @pytest.mark.asyncio |
| 608 | async def test_parse_chat_messages_multiple_images_with_uuids_async( |
| 609 | phi3v_model_config, |
| 610 | image_url, |
| 611 | ): |
| 612 | image_uuid1 = "my_uuid_1" |
| 613 | image_uuid2 = "my_uuid_2" |
| 614 | |
| 615 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 616 | [ |
| 617 | { |
| 618 | "role": "user", |
| 619 | "content": [ |
| 620 | { |
| 621 | "type": "image_url", |
| 622 | "image_url": {"url": image_url}, |
| 623 | "uuid": image_uuid1, |
| 624 | }, |
| 625 | { |
| 626 | "type": "image_pil", |
| 627 | "image_pil": ImageAsset("cherry_blossom").pil_image, |
| 628 | "uuid": image_uuid2, |
| 629 | }, |
| 630 | {"type": "text", "text": "What's in these images?"}, |
| 631 | ], |
| 632 | } |
| 633 | ], |
| 634 | phi3v_model_config, |
| 635 | content_format="string", |
| 636 | ) |
| 637 | |
| 638 | assert conversation == [ |
| 639 | { |
| 640 | "role": "user", |
| 641 | "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?", |
| 642 | } |
| 643 | ] |
| 644 | _assert_mm_data_is_image_input(mm_data, 2) |
| 645 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid1, image_uuid2]) |
| 646 | |
| 647 | |
| 648 | @pytest.mark.asyncio |
| 649 | async def test_parse_chat_messages_multiple_empty_images_with_uuids_async( |
| 650 | phi3v_model_config, |
| 651 | image_url, |
| 652 | ): |
| 653 | image_uuid1 = "my_uuid_1" |
| 654 | image_uuid2 = "my_uuid_2" |
| 655 | |
| 656 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 657 | [ |
| 658 | { |
| 659 | "role": "user", |
| 660 | "content": [ |
| 661 | { |
| 662 | "type": "image_url", |
| 663 | "image_url": None, |
| 664 | "uuid": image_uuid1, |
| 665 | }, |
| 666 | { |
| 667 | "type": "image_pil", |
| 668 | "image_pil": None, |
| 669 | "uuid": image_uuid2, |
| 670 | }, |
| 671 | {"type": "text", "text": "What's in these images?"}, |
| 672 | ], |
| 673 | } |
| 674 | ], |
| 675 | phi3v_model_config, |
| 676 | content_format="string", |
| 677 | ) |
| 678 | |
| 679 | assert conversation == [ |
| 680 | { |
| 681 | "role": "user", |
| 682 | "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?", |
| 683 | } |
| 684 | ] |
| 685 | _assert_mm_data_is_image_input(mm_data, 2, skipped_image_indices=[0, 1]) |
| 686 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid1, image_uuid2]) |
| 687 | |
| 688 | |
| 689 | @pytest.mark.asyncio |
| 690 | async def test_parse_chat_messages_multiple_images_with_partial_uuids_async( |
| 691 | phi3v_model_config, |
| 692 | image_url, |
| 693 | ): |
| 694 | image_uuid2 = "my_uuid_2" |
| 695 | |
| 696 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 697 | [ |
| 698 | { |
| 699 | "role": "user", |
| 700 | "content": [ |
| 701 | { |
| 702 | "type": "image_url", |
| 703 | "image_url": {"url": image_url}, |
| 704 | }, |
| 705 | { |
| 706 | "type": "image_pil", |
| 707 | "image_pil": ImageAsset("cherry_blossom").pil_image, |
| 708 | "uuid": image_uuid2, |
| 709 | }, |
| 710 | {"type": "text", "text": "What's in these images?"}, |
| 711 | ], |
| 712 | } |
| 713 | ], |
| 714 | phi3v_model_config, |
| 715 | content_format="string", |
| 716 | ) |
| 717 | |
| 718 | assert conversation == [ |
| 719 | { |
| 720 | "role": "user", |
| 721 | "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?", |
| 722 | } |
| 723 | ] |
| 724 | _assert_mm_data_is_image_input(mm_data, 2) |
| 725 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, image_uuid2]) |
| 726 | |
| 727 | |
| 728 | def test_parse_chat_messages_empty_system( |
| 729 | mistral_model_config, |
| 730 | ): |
| 731 | # Test string format |
| 732 | conversation, _, _ = parse_chat_messages( |
| 733 | [ |
| 734 | {"role": "system", "content": ""}, |
| 735 | { |
| 736 | "role": "user", |
| 737 | "content": [{"type": "text", "text": "Who are you?"}], |
| 738 | }, |
| 739 | ], |
| 740 | mistral_model_config, |
| 741 | content_format="string", |
| 742 | ) |
| 743 | assert conversation == [ |
| 744 | {"role": "system", "content": ""}, |
| 745 | {"role": "user", "content": "Who are you?"}, |
| 746 | ] |
| 747 | |
| 748 | # Test openai format |
| 749 | conversation, _, _ = parse_chat_messages( |
| 750 | [ |
| 751 | {"role": "system", "content": ""}, |
| 752 | { |
| 753 | "role": "user", |
| 754 | "content": [{"type": "text", "text": "Who are you?"}], |
| 755 | }, |
| 756 | ], |
| 757 | mistral_model_config, |
| 758 | content_format="openai", |
| 759 | ) |
| 760 | assert conversation == [ |
| 761 | {"role": "system", "content": [{"type": "text", "text": ""}]}, |
| 762 | {"role": "user", "content": [{"type": "text", "text": "Who are you?"}]}, |
| 763 | ] |
| 764 | |
| 765 | |
| 766 | @pytest.mark.asyncio |
| 767 | async def test_text_only_chat_does_not_initialize_media_connector( |
| 768 | mistral_model_config, |
| 769 | monkeypatch, |
| 770 | ): |
| 771 | load_connector = MagicMock() |
| 772 | monkeypatch.setattr(MEDIA_CONNECTOR_REGISTRY, "load", load_connector) |
| 773 | messages: list[ChatCompletionMessageParam] = [ |
| 774 | {"role": "user", "content": "Who are you?"} |
| 775 | ] |
| 776 | |
| 777 | parse_chat_messages( |
| 778 | messages, |
| 779 | mistral_model_config, |
| 780 | content_format="string", |
| 781 | ) |
| 782 | await parse_chat_messages_async( |
| 783 | messages, |
| 784 | mistral_model_config, |
| 785 | content_format="string", |
| 786 | ) |
| 787 | |
| 788 | load_connector.assert_not_called() |
| 789 | |
| 790 | |
| 791 | @pytest.mark.asyncio |
| 792 | async def test_parse_chat_messages_single_image_async( |
| 793 | phi3v_model_config, |
| 794 | image_url, |
| 795 | ): |
| 796 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 797 | [ |
| 798 | { |
| 799 | "role": "user", |
| 800 | "content": [ |
| 801 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 802 | {"type": "text", "text": "What's in the image?"}, |
| 803 | ], |
| 804 | } |
| 805 | ], |
| 806 | phi3v_model_config, |
| 807 | content_format="string", |
| 808 | ) |
| 809 | |
| 810 | assert conversation == [ |
| 811 | {"role": "user", "content": "<|image_1|>\nWhat's in the image?"} |
| 812 | ] |
| 813 | _assert_mm_data_is_image_input(mm_data, 1) |
| 814 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None]) |
| 815 | |
| 816 | |
| 817 | def test_parse_chat_messages_multiple_images( |
| 818 | phi3v_model_config, |
| 819 | image_url, |
| 820 | ): |
| 821 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 822 | [ |
| 823 | { |
| 824 | "role": "user", |
| 825 | "content": [ |
| 826 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 827 | { |
| 828 | "type": "image_pil", |
| 829 | "image_pil": ImageAsset("cherry_blossom").pil_image, |
| 830 | }, |
| 831 | {"type": "text", "text": "What's in these images?"}, |
| 832 | ], |
| 833 | } |
| 834 | ], |
| 835 | phi3v_model_config, |
| 836 | content_format="string", |
| 837 | ) |
| 838 | |
| 839 | assert conversation == [ |
| 840 | { |
| 841 | "role": "user", |
| 842 | "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?", |
| 843 | } |
| 844 | ] |
| 845 | _assert_mm_data_is_image_input(mm_data, 2) |
| 846 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None]) |
| 847 | |
| 848 | |
| 849 | def test_parse_chat_messages_empty_pil_image_with_uuid( |
| 850 | phi3v_model_config, |
| 851 | ): |
| 852 | uuid = "abcd" |
| 853 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 854 | [ |
| 855 | { |
| 856 | "role": "user", |
| 857 | "content": [ |
| 858 | {"type": "image_pil", "image_pil": None, "uuid": uuid}, |
| 859 | {"type": "text", "text": "What's in this image?"}, |
| 860 | ], |
| 861 | } |
| 862 | ], |
| 863 | phi3v_model_config, |
| 864 | content_format="string", |
| 865 | ) |
| 866 | |
| 867 | assert conversation == [ |
| 868 | { |
| 869 | "role": "user", |
| 870 | "content": "<|image_1|>\nWhat's in this image?", |
| 871 | } |
| 872 | ] |
| 873 | _assert_mm_data_is_image_input(mm_data, 1, skipped_image_indices=[0]) |
| 874 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[uuid]) |
| 875 | |
| 876 | |
| 877 | def test_parse_chat_messages_empty_image_embeds_with_uuid( |
| 878 | phi3v_model_config_image_embeds, |
| 879 | ): |
| 880 | uuid = "abcd" |
| 881 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 882 | [ |
| 883 | { |
| 884 | "role": "user", |
| 885 | "content": [ |
| 886 | {"type": "image_embeds", "image_embeds": None, "uuid": uuid}, |
| 887 | {"type": "text", "text": "What's in this image?"}, |
| 888 | ], |
| 889 | } |
| 890 | ], |
| 891 | phi3v_model_config_image_embeds, |
| 892 | content_format="string", |
| 893 | ) |
| 894 | |
| 895 | assert conversation == [ |
| 896 | { |
| 897 | "role": "user", |
| 898 | "content": "<|image_1|>\nWhat's in this image?", |
| 899 | } |
| 900 | ] |
| 901 | |
| 902 | assert mm_data is not None |
| 903 | assert "image" in mm_data |
| 904 | assert isinstance(mm_data["image"], list) |
| 905 | assert len(mm_data["image"]) == 1 |
| 906 | assert mm_data["image"][0] is None |
| 907 | |
| 908 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[uuid]) |
| 909 | |
| 910 | |
| 911 | def test_parse_chat_messages_empty_audio_embeds_with_uuid( |
| 912 | audio_embeds_model_config, |
| 913 | ): |
| 914 | """Test audio_embeds with UUID (no actual embeds data).""" |
| 915 | uuid = "test-audio-uuid-123" |
| 916 | |
| 917 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 918 | [ |
| 919 | { |
| 920 | "role": "user", |
| 921 | "content": [ |
| 922 | {"type": "text", "text": "Describe this audio"}, |
| 923 | {"type": "audio_embeds", "audio_embeds": None, "uuid": uuid}, |
| 924 | ], |
| 925 | } |
| 926 | ], |
| 927 | audio_embeds_model_config, |
| 928 | content_format="string", |
| 929 | ) |
| 930 | |
| 931 | # Should have audio in mm_data as None (UUID provided) |
| 932 | assert mm_data is not None |
| 933 | assert "audio" in mm_data |
| 934 | assert isinstance(mm_data["audio"], list) |
| 935 | assert len(mm_data["audio"]) == 1 |
| 936 | assert mm_data["audio"][0] is None |
| 937 | |
| 938 | # UUID should be recorded |
| 939 | _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[uuid]) |
| 940 | |
| 941 | |
| 942 | def test_parse_chat_messages_audio_embeds_with_string( |
| 943 | audio_embeds_model_config, |
| 944 | ): |
| 945 | """Test audio_embeds with base64 string embedding data.""" |
| 946 | import torch |
| 947 | |
| 948 | # Create a sample audio embedding tensor |
| 949 | hidden_size = audio_embeds_model_config.get_inputs_embeds_size() |
| 950 | audio_embedding = torch.randn(1, 128, hidden_size) |
| 951 | |
| 952 | # Encode it as base64 |
| 953 | base64_audio_embedding = tensor2base64(audio_embedding) |
| 954 | |
| 955 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 956 | [ |
| 957 | { |
| 958 | "role": "user", |
| 959 | "content": [ |
| 960 | {"type": "text", "text": "Describe this audio"}, |
| 961 | { |
| 962 | "type": "audio_embeds", |
| 963 | "audio_embeds": base64_audio_embedding, |
| 964 | }, |
| 965 | ], |
| 966 | } |
| 967 | ], |
| 968 | audio_embeds_model_config, |
| 969 | content_format="string", |
| 970 | ) |
| 971 | |
| 972 | # Should have audio embedding in mm_data (single tensor, not a list) |
| 973 | assert mm_data is not None |
| 974 | assert "audio" in mm_data |
| 975 | assert isinstance(mm_data["audio"], torch.Tensor) |
| 976 | assert mm_data["audio"].shape == audio_embedding.shape |
| 977 | # No UUID provided |
| 978 | assert mm_uuids is not None |
| 979 | assert "audio" in mm_uuids |
| 980 | _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[None]) |
| 981 | |
| 982 | |
| 983 | @pytest.mark.asyncio |
| 984 | async def test_parse_chat_messages_audio_embeds_async( |
| 985 | audio_embeds_model_config, |
| 986 | ): |
| 987 | """Test audio_embeds with async futures.""" |
| 988 | import torch |
| 989 | |
| 990 | # Create a sample audio embedding tensor |
| 991 | hidden_size = audio_embeds_model_config.get_inputs_embeds_size() |
| 992 | audio_embedding = torch.randn(1, 128, hidden_size) |
| 993 | |
| 994 | # Encode it as base64 |
| 995 | base64_audio_embedding = tensor2base64(audio_embedding) |
| 996 | |
| 997 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 998 | [ |
| 999 | { |
| 1000 | "role": "user", |
| 1001 | "content": [ |
| 1002 | {"type": "text", "text": "Describe this audio"}, |
| 1003 | { |
| 1004 | "type": "audio_embeds", |
| 1005 | "audio_embeds": base64_audio_embedding, |
| 1006 | }, |
| 1007 | ], |
| 1008 | } |
| 1009 | ], |
| 1010 | audio_embeds_model_config, |
| 1011 | content_format="string", |
| 1012 | ) |
| 1013 | |
| 1014 | # Should have audio embedding in mm_data (single tensor, not a list) |
| 1015 | assert mm_data is not None |
| 1016 | assert "audio" in mm_data |
| 1017 | assert isinstance(mm_data["audio"], torch.Tensor) |
| 1018 | assert mm_data["audio"].shape == audio_embedding.shape |
| 1019 | # No UUID provided |
| 1020 | assert mm_uuids is not None |
| 1021 | assert "audio" in mm_uuids |
| 1022 | _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[None]) |
| 1023 | |
| 1024 | |
| 1025 | def test_parse_chat_messages_empty_video_embeds_with_uuid( |
| 1026 | qwen25omni_model_config_video_embeds, |
| 1027 | ): |
| 1028 | """Test video_embeds with UUID (no actual embeds data).""" |
| 1029 | uuid = "test-video-uuid-123" |
| 1030 | |
| 1031 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1032 | [ |
| 1033 | { |
| 1034 | "role": "user", |
| 1035 | "content": [ |
| 1036 | {"type": "text", "text": "Describe this video"}, |
| 1037 | {"type": "video_embeds", "video_embeds": None, "uuid": uuid}, |
| 1038 | ], |
| 1039 | } |
| 1040 | ], |
| 1041 | qwen25omni_model_config_video_embeds, |
| 1042 | content_format="string", |
| 1043 | ) |
| 1044 | |
| 1045 | # Should have video in mm_data as None (UUID provided) |
| 1046 | assert mm_data is not None |
| 1047 | assert "video" in mm_data |
| 1048 | assert isinstance(mm_data["video"], list) |
| 1049 | assert len(mm_data["video"]) == 1 |
| 1050 | assert mm_data["video"][0] is None |
| 1051 | |
| 1052 | # UUID should be recorded |
| 1053 | _assert_mm_uuids(mm_uuids, 1, modality="video", expected_uuids=[uuid]) |
| 1054 | |
| 1055 | |
| 1056 | def test_parse_chat_messages_video_embeds_with_string( |
| 1057 | qwen25omni_model_config_video_embeds, |
| 1058 | ): |
| 1059 | """Test video_embeds with base64 string embedding data.""" |
| 1060 | hidden_size = qwen25omni_model_config_video_embeds.get_inputs_embeds_size() |
| 1061 | video_embedding = torch.randn(1, 128, hidden_size) |
| 1062 | |
| 1063 | base64_video_embedding = tensor2base64(video_embedding) |
| 1064 | |
| 1065 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1066 | [ |
| 1067 | { |
| 1068 | "role": "user", |
| 1069 | "content": [ |
| 1070 | {"type": "text", "text": "Describe this video"}, |
| 1071 | { |
| 1072 | "type": "video_embeds", |
| 1073 | "video_embeds": base64_video_embedding, |
| 1074 | }, |
| 1075 | ], |
| 1076 | } |
| 1077 | ], |
| 1078 | qwen25omni_model_config_video_embeds, |
| 1079 | content_format="string", |
| 1080 | ) |
| 1081 | |
| 1082 | # Should have video embedding in mm_data (single tensor, not a list) |
| 1083 | assert mm_data is not None |
| 1084 | assert "video" in mm_data |
| 1085 | assert isinstance(mm_data["video"], torch.Tensor) |
| 1086 | assert mm_data["video"].shape == video_embedding.shape |
| 1087 | # No UUID provided |
| 1088 | assert mm_uuids is not None |
| 1089 | assert "video" in mm_uuids |
| 1090 | _assert_mm_uuids(mm_uuids, 1, modality="video", expected_uuids=[None]) |
| 1091 | |
| 1092 | |
| 1093 | @pytest.mark.asyncio |
| 1094 | async def test_parse_chat_messages_video_embeds_async( |
| 1095 | qwen25omni_model_config_video_embeds, |
| 1096 | ): |
| 1097 | """Test video_embeds with async futures.""" |
| 1098 | hidden_size = qwen25omni_model_config_video_embeds.get_inputs_embeds_size() |
| 1099 | video_embedding = torch.randn(1, 128, hidden_size) |
| 1100 | |
| 1101 | base64_video_embedding = tensor2base64(video_embedding) |
| 1102 | |
| 1103 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 1104 | [ |
| 1105 | { |
| 1106 | "role": "user", |
| 1107 | "content": [ |
| 1108 | {"type": "text", "text": "Describe this video"}, |
| 1109 | { |
| 1110 | "type": "video_embeds", |
| 1111 | "video_embeds": base64_video_embedding, |
| 1112 | }, |
| 1113 | ], |
| 1114 | } |
| 1115 | ], |
| 1116 | qwen25omni_model_config_video_embeds, |
| 1117 | content_format="string", |
| 1118 | ) |
| 1119 | |
| 1120 | # Should have video embedding in mm_data (single tensor, not a list) |
| 1121 | assert mm_data is not None |
| 1122 | assert "video" in mm_data |
| 1123 | assert isinstance(mm_data["video"], torch.Tensor) |
| 1124 | assert mm_data["video"].shape == video_embedding.shape |
| 1125 | # No UUID provided |
| 1126 | assert mm_uuids is not None |
| 1127 | assert "video" in mm_uuids |
| 1128 | _assert_mm_uuids(mm_uuids, 1, modality="video", expected_uuids=[None]) |
| 1129 | |
| 1130 | |
| 1131 | def test_parse_chat_messages_video_and_video_embeds_mixed( |
| 1132 | qwen25omni_model_config_video_embeds, |
| 1133 | video_url, |
| 1134 | ): |
| 1135 | """Test that mixing raw video and video_embeds is rejected.""" |
| 1136 | with pytest.raises(VLLMValidationError, match="Mixing raw video"): |
| 1137 | parse_chat_messages( |
| 1138 | [ |
| 1139 | { |
| 1140 | "role": "user", |
| 1141 | "content": [ |
| 1142 | {"type": "video_url", "video_url": {"url": video_url}}, |
| 1143 | {"type": "video_embeds", "video_embeds": None, "uuid": "x"}, |
| 1144 | ], |
| 1145 | } |
| 1146 | ], |
| 1147 | qwen25omni_model_config_video_embeds, |
| 1148 | content_format="string", |
| 1149 | ) |
| 1150 | |
| 1151 | |
| 1152 | def test_parse_chat_messages_multiple_image_embeds( |
| 1153 | phi3v_model_config_image_embeds, |
| 1154 | ): |
| 1155 | """Test that multiple image_embeds in a single message are now supported. |
| 1156 | |
| 1157 | This test validates the fix for the limitation that previously only allowed |
| 1158 | one message with {'type': 'image_embeds'}. Now multiple image embeddings |
| 1159 | can be provided in a single request, similar to regular images. |
| 1160 | """ |
| 1161 | # Create two sample image embedding tensors |
| 1162 | hidden_size = phi3v_model_config_image_embeds.get_inputs_embeds_size() |
| 1163 | image_embedding_1 = torch.randn(256, hidden_size) |
| 1164 | image_embedding_2 = torch.randn(128, hidden_size) |
| 1165 | |
| 1166 | # Encode them as base64 using the convenience function |
| 1167 | base64_image_embedding_1 = tensor2base64(image_embedding_1) |
| 1168 | base64_image_embedding_2 = tensor2base64(image_embedding_2) |
| 1169 | |
| 1170 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1171 | [ |
| 1172 | { |
| 1173 | "role": "user", |
| 1174 | "content": [ |
| 1175 | { |
| 1176 | "type": "image_embeds", |
| 1177 | "image_embeds": base64_image_embedding_1, |
| 1178 | }, |
| 1179 | { |
| 1180 | "type": "image_embeds", |
| 1181 | "image_embeds": base64_image_embedding_2, |
| 1182 | }, |
| 1183 | {"type": "text", "text": "Describe these two images."}, |
| 1184 | ], |
| 1185 | } |
| 1186 | ], |
| 1187 | phi3v_model_config_image_embeds, |
| 1188 | content_format="string", |
| 1189 | ) |
| 1190 | |
| 1191 | # Verify conversation structure |
| 1192 | assert conversation == [ |
| 1193 | { |
| 1194 | "role": "user", |
| 1195 | "content": "<|image_1|>\n<|image_2|>\nDescribe these two images.", |
| 1196 | } |
| 1197 | ] |
| 1198 | |
| 1199 | # Verify mm_data contains a list of embeddings (not a single embedding) |
| 1200 | assert mm_data is not None |
| 1201 | assert "image" in mm_data |
| 1202 | assert isinstance(mm_data["image"], list) |
| 1203 | assert len(mm_data["image"]) == 2 |
| 1204 | |
| 1205 | # Verify each embedding has the correct shape |
| 1206 | assert isinstance(mm_data["image"][0], torch.Tensor) |
| 1207 | assert mm_data["image"][0].shape == image_embedding_1.shape |
| 1208 | assert isinstance(mm_data["image"][1], torch.Tensor) |
| 1209 | assert mm_data["image"][1].shape == image_embedding_2.shape |
| 1210 | |
| 1211 | # Verify UUIDs (None since we didn't provide any) |
| 1212 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None]) |
| 1213 | |
| 1214 | |
| 1215 | def test_parse_chat_messages_multiple_image_embeds_with_uuids( |
| 1216 | phi3v_model_config_image_embeds, |
| 1217 | ): |
| 1218 | """Test multiple image_embeds with UUIDs. |
| 1219 | |
| 1220 | This validates that UUIDs are properly tracked for multiple embeddings. |
| 1221 | """ |
| 1222 | uuid1 = "image-uuid-1" |
| 1223 | uuid2 = "image-uuid-2" |
| 1224 | |
| 1225 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1226 | [ |
| 1227 | { |
| 1228 | "role": "user", |
| 1229 | "content": [ |
| 1230 | { |
| 1231 | "type": "image_embeds", |
| 1232 | "image_embeds": None, |
| 1233 | "uuid": uuid1, |
| 1234 | }, |
| 1235 | { |
| 1236 | "type": "image_embeds", |
| 1237 | "image_embeds": None, |
| 1238 | "uuid": uuid2, |
| 1239 | }, |
| 1240 | {"type": "text", "text": "Compare these images."}, |
| 1241 | ], |
| 1242 | } |
| 1243 | ], |
| 1244 | phi3v_model_config_image_embeds, |
| 1245 | content_format="string", |
| 1246 | ) |
| 1247 | |
| 1248 | # Verify conversation structure |
| 1249 | assert conversation == [ |
| 1250 | { |
| 1251 | "role": "user", |
| 1252 | "content": "<|image_1|>\n<|image_2|>\nCompare these images.", |
| 1253 | } |
| 1254 | ] |
| 1255 | |
| 1256 | # Verify mm_data contains a list with None values (UUID references) |
| 1257 | assert mm_data is not None |
| 1258 | assert "image" in mm_data |
| 1259 | assert isinstance(mm_data["image"], list) |
| 1260 | assert len(mm_data["image"]) == 2 |
| 1261 | assert mm_data["image"][0] is None |
| 1262 | assert mm_data["image"][1] is None |
| 1263 | |
| 1264 | # Verify UUIDs are correctly tracked |
| 1265 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[uuid1, uuid2]) |
| 1266 | |
| 1267 | |
| 1268 | @pytest.mark.asyncio |
| 1269 | async def test_parse_chat_messages_multiple_image_embeds_async( |
| 1270 | phi3v_model_config_image_embeds, |
| 1271 | ): |
| 1272 | """Test multiple image_embeds with async parsing. |
| 1273 | |
| 1274 | This validates the AsyncMultiModalItemTracker also supports multiple embeddings. |
| 1275 | """ |
| 1276 | # Create two sample image embedding tensors |
| 1277 | hidden_size = phi3v_model_config_image_embeds.get_inputs_embeds_size() |
| 1278 | image_embedding_1 = torch.randn(200, hidden_size) |
| 1279 | image_embedding_2 = torch.randn(150, hidden_size) |
| 1280 | |
| 1281 | # Encode them as base64 using the convenience function |
| 1282 | base64_image_embedding_1 = tensor2base64(image_embedding_1) |
| 1283 | base64_image_embedding_2 = tensor2base64(image_embedding_2) |
| 1284 | |
| 1285 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 1286 | [ |
| 1287 | { |
| 1288 | "role": "user", |
| 1289 | "content": [ |
| 1290 | { |
| 1291 | "type": "image_embeds", |
| 1292 | "image_embeds": base64_image_embedding_1, |
| 1293 | }, |
| 1294 | { |
| 1295 | "type": "image_embeds", |
| 1296 | "image_embeds": base64_image_embedding_2, |
| 1297 | }, |
| 1298 | {"type": "text", "text": "What do these images show?"}, |
| 1299 | ], |
| 1300 | } |
| 1301 | ], |
| 1302 | phi3v_model_config_image_embeds, |
| 1303 | content_format="string", |
| 1304 | ) |
| 1305 | |
| 1306 | # Verify conversation structure |
| 1307 | assert conversation == [ |
| 1308 | { |
| 1309 | "role": "user", |
| 1310 | "content": "<|image_1|>\n<|image_2|>\nWhat do these images show?", |
| 1311 | } |
| 1312 | ] |
| 1313 | |
| 1314 | # Await the future and verify mm_data |
| 1315 | assert mm_data is not None |
| 1316 | assert "image" in mm_data |
| 1317 | assert isinstance(mm_data["image"], list) |
| 1318 | assert len(mm_data["image"]) == 2 |
| 1319 | |
| 1320 | # Verify each embedding has the correct shape |
| 1321 | assert isinstance(mm_data["image"][0], torch.Tensor) |
| 1322 | assert mm_data["image"][0].shape == image_embedding_1.shape |
| 1323 | assert isinstance(mm_data["image"][1], torch.Tensor) |
| 1324 | assert mm_data["image"][1].shape == image_embedding_2.shape |
| 1325 | |
| 1326 | # Verify UUIDs |
| 1327 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None]) |
| 1328 | |
| 1329 | |
| 1330 | @pytest.mark.asyncio |
| 1331 | async def test_parse_chat_messages_empty_image_embeds_with_uuid_async( |
| 1332 | phi3v_model_config_image_embeds, |
| 1333 | ): |
| 1334 | uuid = "abcd" |
| 1335 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 1336 | [ |
| 1337 | { |
| 1338 | "role": "user", |
| 1339 | "content": [ |
| 1340 | {"type": "image_embeds", "image_embeds": None, "uuid": uuid}, |
| 1341 | {"type": "text", "text": "What's in this image?"}, |
| 1342 | ], |
| 1343 | } |
| 1344 | ], |
| 1345 | phi3v_model_config_image_embeds, |
| 1346 | content_format="string", |
| 1347 | ) |
| 1348 | |
| 1349 | assert conversation == [ |
| 1350 | { |
| 1351 | "role": "user", |
| 1352 | "content": "<|image_1|>\nWhat's in this image?", |
| 1353 | } |
| 1354 | ] |
| 1355 | assert mm_data is not None |
| 1356 | assert "image" in mm_data |
| 1357 | assert isinstance(mm_data["image"], list) |
| 1358 | assert len(mm_data["image"]) == 1 |
| 1359 | assert mm_data["image"][0] is None |
| 1360 | |
| 1361 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[uuid]) |
| 1362 | |
| 1363 | |
| 1364 | def test_parse_chat_messages_empty_dict_image_embeds( |
| 1365 | phi3v_model_config_image_embeds, |
| 1366 | ): |
| 1367 | """Test that empty dictionary for image_embeds is handled without errors.""" |
| 1368 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1369 | [ |
| 1370 | { |
| 1371 | "role": "user", |
| 1372 | "content": [ |
| 1373 | {"type": "image_embeds", "image_embeds": {}}, |
| 1374 | {"type": "text", "text": "What's in this image?"}, |
| 1375 | ], |
| 1376 | } |
| 1377 | ], |
| 1378 | phi3v_model_config_image_embeds, |
| 1379 | content_format="string", |
| 1380 | ) |
| 1381 | |
| 1382 | # Verify conversation structure |
| 1383 | assert conversation == [ |
| 1384 | { |
| 1385 | "role": "user", |
| 1386 | "content": "<|image_1|>\nWhat's in this image?", |
| 1387 | } |
| 1388 | ] |
| 1389 | |
| 1390 | # Verify mm_data contains an empty dictionary of embeddings |
| 1391 | assert mm_data is not None |
| 1392 | assert "image" in mm_data |
| 1393 | assert isinstance(mm_data["image"], dict) |
| 1394 | assert len(mm_data["image"]) == 0 |
| 1395 | |
| 1396 | # Verify UUIDs (None since we didn't provide any) |
| 1397 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None]) |
| 1398 | |
| 1399 | |
| 1400 | @pytest.mark.parametrize("json_metadata", [False, True]) |
| 1401 | def test_parse_chat_messages_multiple_dict_image_embeds( |
| 1402 | qwen25omni_model_config_image_embeds, |
| 1403 | json_metadata, |
| 1404 | ): |
| 1405 | """Test that multiple dictionaries for image_embeds is handled without errors.""" |
| 1406 | # Create two sample image embedding tensors |
| 1407 | batch_size = 2 |
| 1408 | hidden_size = qwen25omni_model_config_image_embeds.get_inputs_embeds_size() |
| 1409 | image_embeds = torch.randn(batch_size * 220, hidden_size) |
| 1410 | image_grid_thw = torch.tensor([[1, 22, 40] for _ in range(batch_size)]) |
| 1411 | |
| 1412 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1413 | [ |
| 1414 | { |
| 1415 | "role": "user", |
| 1416 | "content": [ |
| 1417 | { |
| 1418 | "type": "image_embeds", |
| 1419 | "image_embeds": { |
| 1420 | "image_embeds": tensor2base64(embeds), |
| 1421 | "image_grid_thw": grid_thw.tolist() |
| 1422 | if json_metadata |
| 1423 | else tensor2base64(grid_thw), |
| 1424 | }, |
| 1425 | } |
| 1426 | for embeds, grid_thw in zip( |
| 1427 | image_embeds.chunk(batch_size), image_grid_thw |
| 1428 | ) |
| 1429 | ] |
| 1430 | + [ |
| 1431 | {"type": "text", "text": "Describe these two images."}, |
| 1432 | ], |
| 1433 | } |
| 1434 | ], |
| 1435 | qwen25omni_model_config_image_embeds, |
| 1436 | content_format="string", |
| 1437 | ) |
| 1438 | |
| 1439 | # Verify conversation structure |
| 1440 | assert conversation == [ |
| 1441 | { |
| 1442 | "role": "user", |
| 1443 | "content": "<|vision_start|><|IMAGE|><|vision_end|>\n" |
| 1444 | "<|vision_start|><|IMAGE|><|vision_end|>\nDescribe these two images.", |
| 1445 | } |
| 1446 | ] |
| 1447 | |
| 1448 | # Verify mm_data contains a dictionary of multi-embeddings |
| 1449 | assert mm_data is not None |
| 1450 | assert "image" in mm_data |
| 1451 | assert isinstance(mm_data["image"], dict) |
| 1452 | assert len(mm_data["image"]) == batch_size |
| 1453 | |
| 1454 | # Verify each embedding has the correct shape |
| 1455 | assert isinstance(mm_data["image"]["image_embeds"], torch.Tensor) |
| 1456 | assert mm_data["image"]["image_embeds"].shape == image_embeds.shape |
| 1457 | assert isinstance(mm_data["image"]["image_grid_thw"], torch.Tensor) |
| 1458 | assert mm_data["image"]["image_grid_thw"].shape == image_grid_thw.shape |
| 1459 | |
| 1460 | # Verify UUIDs (None since we didn't provide any) |
| 1461 | _assert_mm_uuids(mm_uuids, batch_size, expected_uuids=[None, None]) |
| 1462 | |
| 1463 | |
| 1464 | @pytest.mark.asyncio |
| 1465 | async def test_parse_chat_messages_multiple_images_async( |
| 1466 | phi3v_model_config, |
| 1467 | image_url, |
| 1468 | ): |
| 1469 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 1470 | [ |
| 1471 | { |
| 1472 | "role": "user", |
| 1473 | "content": [ |
| 1474 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1475 | { |
| 1476 | "type": "image_pil", |
| 1477 | "image_pil": ImageAsset("cherry_blossom").pil_image, |
| 1478 | }, |
| 1479 | {"type": "text", "text": "What's in these images?"}, |
| 1480 | ], |
| 1481 | } |
| 1482 | ], |
| 1483 | phi3v_model_config, |
| 1484 | content_format="string", |
| 1485 | ) |
| 1486 | |
| 1487 | assert conversation == [ |
| 1488 | { |
| 1489 | "role": "user", |
| 1490 | "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?", |
| 1491 | } |
| 1492 | ] |
| 1493 | _assert_mm_data_is_image_input(mm_data, 2) |
| 1494 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None]) |
| 1495 | |
| 1496 | |
| 1497 | def test_parse_chat_messages_placeholder_already_in_prompt( |
| 1498 | phi3v_model_config, |
| 1499 | image_url, |
| 1500 | ): |
| 1501 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1502 | [ |
| 1503 | { |
| 1504 | "role": "user", |
| 1505 | "content": [ |
| 1506 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1507 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1508 | { |
| 1509 | "type": "text", |
| 1510 | "text": "What's in <|image_1|> and how does it compare to <|image_2|>?", # noqa: E501 |
| 1511 | }, |
| 1512 | ], |
| 1513 | } |
| 1514 | ], |
| 1515 | phi3v_model_config, |
| 1516 | content_format="string", |
| 1517 | ) |
| 1518 | assert conversation == [ |
| 1519 | { |
| 1520 | "role": "user", |
| 1521 | "content": "What's in <|image_1|> and how does it compare to <|image_2|>?", |
| 1522 | } |
| 1523 | ] |
| 1524 | _assert_mm_data_is_image_input(mm_data, 2) |
| 1525 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None]) |
| 1526 | |
| 1527 | |
| 1528 | def test_parse_chat_messages_placeholder_one_already_in_prompt( |
| 1529 | phi3v_model_config, |
| 1530 | image_url, |
| 1531 | ): |
| 1532 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1533 | [ |
| 1534 | { |
| 1535 | "role": "user", |
| 1536 | "content": [ |
| 1537 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1538 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1539 | { |
| 1540 | "type": "text", |
| 1541 | "text": "What's in <|image_1|> and how does it compare to " |
| 1542 | "the other one?", |
| 1543 | }, |
| 1544 | ], |
| 1545 | } |
| 1546 | ], |
| 1547 | phi3v_model_config, |
| 1548 | content_format="string", |
| 1549 | ) |
| 1550 | |
| 1551 | assert conversation == [ |
| 1552 | { |
| 1553 | "role": "user", |
| 1554 | "content": "<|image_2|>\nWhat's in <|image_1|> and how does it compare to " |
| 1555 | "the other one?", |
| 1556 | } |
| 1557 | ] |
| 1558 | _assert_mm_data_is_image_input(mm_data, 2) |
| 1559 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None]) |
| 1560 | |
| 1561 | |
| 1562 | def test_parse_chat_messages_multiple_images_across_messages( |
| 1563 | phi3v_model_config, |
| 1564 | image_url, |
| 1565 | ): |
| 1566 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1567 | [ |
| 1568 | { |
| 1569 | "role": "user", |
| 1570 | "content": [ |
| 1571 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1572 | {"type": "text", "text": "What's in this image?"}, |
| 1573 | ], |
| 1574 | }, |
| 1575 | {"role": "assistant", "content": "Some stuff."}, |
| 1576 | { |
| 1577 | "role": "user", |
| 1578 | "content": [ |
| 1579 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1580 | {"type": "text", "text": "What about this one?"}, |
| 1581 | ], |
| 1582 | }, |
| 1583 | ], |
| 1584 | phi3v_model_config, |
| 1585 | content_format="string", |
| 1586 | ) |
| 1587 | |
| 1588 | assert conversation == [ |
| 1589 | {"role": "user", "content": "<|image_1|>\nWhat's in this image?"}, |
| 1590 | {"role": "assistant", "content": "Some stuff."}, |
| 1591 | {"role": "user", "content": "<|image_2|>\nWhat about this one?"}, |
| 1592 | ] |
| 1593 | _assert_mm_data_is_image_input(mm_data, 2) |
| 1594 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None]) |
| 1595 | |
| 1596 | |
| 1597 | def test_parse_chat_messages_multiple_images_with_uuids_across_messages( |
| 1598 | phi3v_model_config, |
| 1599 | image_url, |
| 1600 | ): |
| 1601 | image_uuid = str(hash(image_url)) |
| 1602 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1603 | [ |
| 1604 | { |
| 1605 | "role": "user", |
| 1606 | "content": [ |
| 1607 | { |
| 1608 | "type": "image_url", |
| 1609 | "image_url": {"url": image_url}, |
| 1610 | "uuid": image_uuid, |
| 1611 | }, |
| 1612 | {"type": "text", "text": "What's in this image?"}, |
| 1613 | ], |
| 1614 | }, |
| 1615 | {"role": "assistant", "content": "Some stuff."}, |
| 1616 | { |
| 1617 | "role": "user", |
| 1618 | "content": [ |
| 1619 | { |
| 1620 | "type": "image_url", |
| 1621 | "image_url": {"url": image_url}, |
| 1622 | "uuid": image_uuid, |
| 1623 | }, |
| 1624 | {"type": "text", "text": "What about this one?"}, |
| 1625 | ], |
| 1626 | }, |
| 1627 | ], |
| 1628 | phi3v_model_config, |
| 1629 | content_format="string", |
| 1630 | ) |
| 1631 | |
| 1632 | assert conversation == [ |
| 1633 | {"role": "user", "content": "<|image_1|>\nWhat's in this image?"}, |
| 1634 | {"role": "assistant", "content": "Some stuff."}, |
| 1635 | {"role": "user", "content": "<|image_2|>\nWhat about this one?"}, |
| 1636 | ] |
| 1637 | _assert_mm_data_is_image_input(mm_data, 2) |
| 1638 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid, image_uuid]) |
| 1639 | |
| 1640 | |
| 1641 | def test_parse_chat_messages_context_text_format( |
| 1642 | phi3v_model_config, |
| 1643 | ): |
| 1644 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1645 | [ |
| 1646 | { |
| 1647 | "role": "user", |
| 1648 | "content": [{"type": "text", "text": "What's in this text?"}], |
| 1649 | }, |
| 1650 | {"role": "assistant", "content": "Some stuff."}, |
| 1651 | {"role": "user", "content": "What about this one?"}, |
| 1652 | ], |
| 1653 | phi3v_model_config, |
| 1654 | content_format="openai", |
| 1655 | ) |
| 1656 | |
| 1657 | assert conversation == [ |
| 1658 | { |
| 1659 | "role": "user", |
| 1660 | "content": [{"type": "text", "text": "What's in this text?"}], |
| 1661 | }, |
| 1662 | { |
| 1663 | "role": "assistant", |
| 1664 | "content": [{"type": "text", "text": "Some stuff."}], |
| 1665 | }, |
| 1666 | { |
| 1667 | "role": "user", |
| 1668 | "content": [{"type": "text", "text": "What about this one?"}], |
| 1669 | }, |
| 1670 | ] |
| 1671 | assert mm_data is None |
| 1672 | assert mm_uuids is None |
| 1673 | |
| 1674 | |
| 1675 | def test_parse_chat_messages_openai_format_image_url( |
| 1676 | phi3v_model_config, |
| 1677 | image_url, |
| 1678 | ): |
| 1679 | content = [ |
| 1680 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1681 | {"type": "text", "text": "What's in the image?"}, |
| 1682 | ] |
| 1683 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1684 | [ |
| 1685 | { |
| 1686 | "role": "user", |
| 1687 | "content": content, |
| 1688 | } |
| 1689 | ], |
| 1690 | phi3v_model_config, |
| 1691 | content_format="openai", |
| 1692 | ) |
| 1693 | |
| 1694 | assert conversation == [ |
| 1695 | { |
| 1696 | "role": "user", |
| 1697 | "content": [ |
| 1698 | {"type": "image"}, |
| 1699 | {"type": "text", "text": "What's in the image?"}, |
| 1700 | ], |
| 1701 | } |
| 1702 | ] |
| 1703 | _assert_mm_data_is_image_input(mm_data, 1) |
| 1704 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None]) |
| 1705 | |
| 1706 | |
| 1707 | def test_parse_chat_messages_rejects_too_many_images_in_one_message( |
| 1708 | phi3v_model_config, |
| 1709 | image_url, |
| 1710 | ): |
| 1711 | with warnings.catch_warnings(): |
| 1712 | warnings.filterwarnings( |
| 1713 | "ignore", |
| 1714 | message="coroutine 'async_get_and_parse_image' was never awaited", |
| 1715 | ) |
| 1716 | with pytest.raises(VLLMValidationError, match="At most"): |
| 1717 | parse_chat_messages( |
| 1718 | [ |
| 1719 | { |
| 1720 | "role": "user", |
| 1721 | "content": [ |
| 1722 | { |
| 1723 | "type": "image_url", |
| 1724 | "image_url": {"url": image_url}, |
| 1725 | }, |
| 1726 | { |
| 1727 | "type": "image_url", |
| 1728 | "image_url": {"url": image_url}, |
| 1729 | }, |
| 1730 | { |
| 1731 | "type": "image_url", |
| 1732 | "image_url": {"url": image_url}, |
| 1733 | }, |
| 1734 | {"type": "text", "text": "What's in these images?"}, |
| 1735 | ], |
| 1736 | } |
| 1737 | ], |
| 1738 | phi3v_model_config, |
| 1739 | content_format="string", |
| 1740 | ) |
| 1741 | |
| 1742 | |
| 1743 | def test_parse_chat_messages_rejects_too_many_images_across_messages( |
| 1744 | phi3v_model_config, |
| 1745 | image_url, |
| 1746 | ): |
| 1747 | with warnings.catch_warnings(): |
| 1748 | warnings.filterwarnings( |
| 1749 | "ignore", |
| 1750 | message="coroutine 'async_get_and_parse_image' was never awaited", |
| 1751 | ) |
| 1752 | with pytest.raises(VLLMValidationError, match="At most"): |
| 1753 | parse_chat_messages( |
| 1754 | [ |
| 1755 | { |
| 1756 | "role": "user", |
| 1757 | "content": [ |
| 1758 | { |
| 1759 | "type": "image_url", |
| 1760 | "image_url": {"url": image_url}, |
| 1761 | }, |
| 1762 | {"type": "text", "text": "What's in this image?"}, |
| 1763 | ], |
| 1764 | }, |
| 1765 | {"role": "assistant", "content": "Some stuff."}, |
| 1766 | { |
| 1767 | "role": "user", |
| 1768 | "content": [ |
| 1769 | { |
| 1770 | "type": "image_url", |
| 1771 | "image_url": {"url": image_url}, |
| 1772 | }, |
| 1773 | { |
| 1774 | "type": "image_url", |
| 1775 | "image_url": {"url": image_url}, |
| 1776 | }, |
| 1777 | {"type": "text", "text": "What about these two?"}, |
| 1778 | ], |
| 1779 | }, |
| 1780 | ], |
| 1781 | phi3v_model_config, |
| 1782 | content_format="string", |
| 1783 | ) |
| 1784 | |
| 1785 | |
| 1786 | def test_parse_chat_messages_multiple_images_uncommon_input( |
| 1787 | phi3v_model_config, |
| 1788 | image_url, |
| 1789 | ): |
| 1790 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1791 | [ |
| 1792 | { |
| 1793 | "role": "user", |
| 1794 | "content": [ |
| 1795 | "What's in these images?", |
| 1796 | {"image_url": image_url}, |
| 1797 | {"image_url": image_url}, |
| 1798 | ], |
| 1799 | } |
| 1800 | ], |
| 1801 | phi3v_model_config, |
| 1802 | content_format="string", |
| 1803 | ) |
| 1804 | |
| 1805 | assert conversation == [ |
| 1806 | { |
| 1807 | "role": "user", |
| 1808 | "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?", |
| 1809 | } |
| 1810 | ] |
| 1811 | _assert_mm_data_is_image_input(mm_data, 2) |
| 1812 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None]) |
| 1813 | |
| 1814 | |
| 1815 | def test_parse_chat_messages_multiple_images_interleave( |
| 1816 | phi3v_model_config_mm_interleaved, |
| 1817 | image_url, |
| 1818 | ): |
| 1819 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1820 | [ |
| 1821 | { |
| 1822 | "role": "user", |
| 1823 | "content": [ |
| 1824 | { |
| 1825 | "type": "text", |
| 1826 | "text": "I need you to compare this image", |
| 1827 | }, |
| 1828 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1829 | {"type": "text", "text": "and this one"}, |
| 1830 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1831 | {"type": "text", "text": "Do they have differences?"}, |
| 1832 | ], |
| 1833 | } |
| 1834 | ], |
| 1835 | phi3v_model_config_mm_interleaved, |
| 1836 | content_format="string", |
| 1837 | ) |
| 1838 | |
| 1839 | assert conversation == [ |
| 1840 | { |
| 1841 | "role": "user", |
| 1842 | "content": "I need you to compare this image\n<|image_1|>\nand this one\n<|image_2|>\n" # noqa: E501 |
| 1843 | "Do they have differences?", |
| 1844 | } |
| 1845 | ] |
| 1846 | _assert_mm_data_is_image_input(mm_data, 2) |
| 1847 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None]) |
| 1848 | |
| 1849 | |
| 1850 | def test_parse_chat_messages_interleave_placeholder_overcount_raises( |
| 1851 | phi3v_model_config_mm_interleaved, |
| 1852 | image_url, |
| 1853 | ): |
| 1854 | # Only one image is supplied, but the user's text also contains the internal |
| 1855 | # image sentinel, so the interleaved prompt references more image |
| 1856 | # placeholders than actual images. This must surface as a clear validation |
| 1857 | # error rather than an IndexError from popping an empty list. |
| 1858 | with pytest.raises(VLLMValidationError, match="placeholders in input prompt"): |
| 1859 | parse_chat_messages( |
| 1860 | [ |
| 1861 | { |
| 1862 | "role": "user", |
| 1863 | "content": [ |
| 1864 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1865 | {"type": "text", "text": "<##IMAGE##>"}, |
| 1866 | ], |
| 1867 | } |
| 1868 | ], |
| 1869 | phi3v_model_config_mm_interleaved, |
| 1870 | content_format="string", |
| 1871 | ) |
| 1872 | |
| 1873 | |
| 1874 | @pytest.mark.asyncio |
| 1875 | async def test_parse_chat_messages_multiple_images_interleave_async( |
| 1876 | phi3v_model_config_mm_interleaved, |
| 1877 | image_url, |
| 1878 | ): |
| 1879 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 1880 | [ |
| 1881 | { |
| 1882 | "role": "user", |
| 1883 | "content": [ |
| 1884 | { |
| 1885 | "type": "text", |
| 1886 | "text": "I need you to compare this image", |
| 1887 | }, |
| 1888 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1889 | {"type": "text", "text": "and this one"}, |
| 1890 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1891 | {"type": "text", "text": "Do they have differences?"}, |
| 1892 | ], |
| 1893 | } |
| 1894 | ], |
| 1895 | phi3v_model_config_mm_interleaved, |
| 1896 | content_format="string", |
| 1897 | ) |
| 1898 | |
| 1899 | assert conversation == [ |
| 1900 | { |
| 1901 | "role": "user", |
| 1902 | "content": "I need you to compare this image\n<|image_1|>\nand this one\n<|image_2|>\n" # noqa: E501 |
| 1903 | "Do they have differences?", |
| 1904 | } |
| 1905 | ] |
| 1906 | _assert_mm_data_is_image_input(mm_data, 2) |
| 1907 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None]) |
| 1908 | |
| 1909 | |
| 1910 | @pytest.mark.asyncio |
| 1911 | async def test_parse_chat_messages_multiple_images_with_uuids_interleave_async( |
| 1912 | phi3v_model_config_mm_interleaved, |
| 1913 | image_url, |
| 1914 | ): |
| 1915 | image_uuid = str(hash(image_url)) |
| 1916 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 1917 | [ |
| 1918 | { |
| 1919 | "role": "user", |
| 1920 | "content": [ |
| 1921 | { |
| 1922 | "type": "text", |
| 1923 | "text": "I need you to compare this image", |
| 1924 | }, |
| 1925 | { |
| 1926 | "type": "image_url", |
| 1927 | "image_url": {"url": image_url}, |
| 1928 | "uuid": image_uuid, |
| 1929 | }, |
| 1930 | {"type": "text", "text": "and this one"}, |
| 1931 | { |
| 1932 | "type": "image_url", |
| 1933 | "image_url": {"url": image_url}, |
| 1934 | "uuid": image_uuid, |
| 1935 | }, |
| 1936 | {"type": "text", "text": "Do they have differences?"}, |
| 1937 | ], |
| 1938 | } |
| 1939 | ], |
| 1940 | phi3v_model_config_mm_interleaved, |
| 1941 | content_format="string", |
| 1942 | ) |
| 1943 | |
| 1944 | assert conversation == [ |
| 1945 | { |
| 1946 | "role": "user", |
| 1947 | "content": "I need you to compare this image\n<|image_1|>\nand this one\n<|image_2|>\n" # noqa: E501 |
| 1948 | "Do they have differences?", |
| 1949 | } |
| 1950 | ] |
| 1951 | _assert_mm_data_is_image_input(mm_data, 2) |
| 1952 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid, image_uuid]) |
| 1953 | |
| 1954 | |
| 1955 | def test_parse_chat_messages_multiple_images_multiple_messages_interleave( |
| 1956 | phi3v_model_config_mm_interleaved, |
| 1957 | image_url, |
| 1958 | ): |
| 1959 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 1960 | [ |
| 1961 | { |
| 1962 | "role": "user", |
| 1963 | "content": [ |
| 1964 | {"type": "text", "text": "What's on this image?"}, |
| 1965 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1966 | {"type": "text", "text": "Be accurate."}, |
| 1967 | ], |
| 1968 | }, |
| 1969 | {"role": "assistant", "content": "Some stuff."}, |
| 1970 | { |
| 1971 | "role": "user", |
| 1972 | "content": [ |
| 1973 | {"type": "text", "text": "What's on this image?"}, |
| 1974 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 1975 | ], |
| 1976 | }, |
| 1977 | ], |
| 1978 | phi3v_model_config_mm_interleaved, |
| 1979 | content_format="string", |
| 1980 | ) |
| 1981 | |
| 1982 | assert conversation == [ |
| 1983 | { |
| 1984 | "role": "user", |
| 1985 | "content": "What's on this image?\n<|image_1|>\nBe accurate.", |
| 1986 | }, |
| 1987 | {"role": "assistant", "content": "Some stuff."}, |
| 1988 | {"role": "user", "content": "What's on this image?\n<|image_2|>"}, |
| 1989 | ] |
| 1990 | _assert_mm_data_is_image_input(mm_data, 2) |
| 1991 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None]) |
| 1992 | |
| 1993 | |
| 1994 | def test_parse_chat_messages_multiple_images_with_uuids_multiple_messages_interleave( |
| 1995 | phi3v_model_config_mm_interleaved, |
| 1996 | image_url, |
| 1997 | ): |
| 1998 | image_uuid = str(hash(image_url)) |
| 1999 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2000 | [ |
| 2001 | { |
| 2002 | "role": "user", |
| 2003 | "content": [ |
| 2004 | {"type": "text", "text": "What's on this image?"}, |
| 2005 | { |
| 2006 | "type": "image_url", |
| 2007 | "image_url": {"url": image_url}, |
| 2008 | "uuid": image_uuid, |
| 2009 | }, |
| 2010 | {"type": "text", "text": "Be accurate."}, |
| 2011 | ], |
| 2012 | }, |
| 2013 | {"role": "assistant", "content": "Some stuff."}, |
| 2014 | { |
| 2015 | "role": "user", |
| 2016 | "content": [ |
| 2017 | {"type": "text", "text": "What's on this image?"}, |
| 2018 | { |
| 2019 | "type": "image_url", |
| 2020 | "image_url": {"url": image_url}, |
| 2021 | "uuid": image_uuid, |
| 2022 | }, |
| 2023 | ], |
| 2024 | }, |
| 2025 | ], |
| 2026 | phi3v_model_config_mm_interleaved, |
| 2027 | content_format="string", |
| 2028 | ) |
| 2029 | |
| 2030 | assert conversation == [ |
| 2031 | { |
| 2032 | "role": "user", |
| 2033 | "content": "What's on this image?\n<|image_1|>\nBe accurate.", |
| 2034 | }, |
| 2035 | {"role": "assistant", "content": "Some stuff."}, |
| 2036 | {"role": "user", "content": "What's on this image?\n<|image_2|>"}, |
| 2037 | ] |
| 2038 | _assert_mm_data_is_image_input(mm_data, 2) |
| 2039 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid, image_uuid]) |
| 2040 | |
| 2041 | |
| 2042 | def test_parse_chat_messages_multiple_modals_multiple_messages_interleave( |
| 2043 | qwen25omni_model_config_mm_interleaved, |
| 2044 | image_url, |
| 2045 | video_url, |
| 2046 | audio_url, |
| 2047 | ): |
| 2048 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2049 | [ |
| 2050 | { |
| 2051 | "role": "user", |
| 2052 | "content": [ |
| 2053 | {"type": "text", "text": "What's on this image?"}, |
| 2054 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 2055 | {"type": "text", "text": "Now listen to this audio"}, |
| 2056 | {"type": "audio_url", "audio_url": {"url": audio_url}}, |
| 2057 | ], |
| 2058 | }, |
| 2059 | {"role": "assistant", "content": "Some stuff."}, |
| 2060 | { |
| 2061 | "role": "user", |
| 2062 | "content": [ |
| 2063 | {"type": "text", "text": "What's on this image?"}, |
| 2064 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 2065 | {"type": "text", "text": "And what's in the video?"}, |
| 2066 | {"type": "video_url", "video_url": {"url": video_url}}, |
| 2067 | ], |
| 2068 | }, |
| 2069 | ], |
| 2070 | qwen25omni_model_config_mm_interleaved, |
| 2071 | content_format="string", |
| 2072 | ) |
| 2073 | |
| 2074 | assert conversation == [ |
| 2075 | { |
| 2076 | "role": "user", |
| 2077 | "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>" |
| 2078 | "\nNow listen to this audio\nAudio 1: <|audio_bos|><|AUDIO|><|audio_eos|>", |
| 2079 | }, |
| 2080 | {"role": "assistant", "content": "Some stuff."}, |
| 2081 | { |
| 2082 | "role": "user", |
| 2083 | "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>" |
| 2084 | "\nAnd what's in the video?\n<|vision_start|><|VIDEO|><|vision_end|>", |
| 2085 | }, |
| 2086 | ] |
| 2087 | |
| 2088 | _assert_mm_data_inputs(mm_data, {"image": 2, "video": 1, "audio": 1}) |
| 2089 | _assert_mm_uuids(mm_uuids, 2, modality="image", expected_uuids=[None, None]) |
| 2090 | _assert_mm_uuids(mm_uuids, 1, modality="video", expected_uuids=[None]) |
| 2091 | _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[None]) |
| 2092 | |
| 2093 | |
| 2094 | def test_parse_chat_messages_multiple_modals_with_uuids_multiple_messages_interleave( |
| 2095 | qwen25omni_model_config_mm_interleaved, |
| 2096 | image_url, |
| 2097 | video_url, |
| 2098 | audio_url, |
| 2099 | ): |
| 2100 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2101 | [ |
| 2102 | { |
| 2103 | "role": "user", |
| 2104 | "content": [ |
| 2105 | {"type": "text", "text": "What's on this image?"}, |
| 2106 | { |
| 2107 | "type": "image_url", |
| 2108 | "image_url": {"url": image_url}, |
| 2109 | "uuid": "image_123", |
| 2110 | }, |
| 2111 | {"type": "text", "text": "Now listen to this audio"}, |
| 2112 | { |
| 2113 | "type": "audio_url", |
| 2114 | "audio_url": {"url": audio_url}, |
| 2115 | "uuid": "audio_123", |
| 2116 | }, |
| 2117 | ], |
| 2118 | }, |
| 2119 | {"role": "assistant", "content": "Some stuff."}, |
| 2120 | { |
| 2121 | "role": "user", |
| 2122 | "content": [ |
| 2123 | {"type": "text", "text": "What's on this image?"}, |
| 2124 | { |
| 2125 | "type": "image_url", |
| 2126 | "image_url": {"url": image_url}, |
| 2127 | "uuid": "image_123", |
| 2128 | }, |
| 2129 | {"type": "text", "text": "And what's in the video?"}, |
| 2130 | { |
| 2131 | "type": "video_url", |
| 2132 | "video_url": {"url": video_url}, |
| 2133 | "uuid": "video_123", |
| 2134 | }, |
| 2135 | ], |
| 2136 | }, |
| 2137 | ], |
| 2138 | qwen25omni_model_config_mm_interleaved, |
| 2139 | content_format="string", |
| 2140 | ) |
| 2141 | |
| 2142 | assert conversation == [ |
| 2143 | { |
| 2144 | "role": "user", |
| 2145 | "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>" |
| 2146 | "\nNow listen to this audio\nAudio 1: <|audio_bos|><|AUDIO|><|audio_eos|>", |
| 2147 | }, |
| 2148 | {"role": "assistant", "content": "Some stuff."}, |
| 2149 | { |
| 2150 | "role": "user", |
| 2151 | "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>" |
| 2152 | "\nAnd what's in the video?\n<|vision_start|><|VIDEO|><|vision_end|>", |
| 2153 | }, |
| 2154 | ] |
| 2155 | |
| 2156 | _assert_mm_data_inputs(mm_data, {"image": 2, "video": 1, "audio": 1}) |
| 2157 | _assert_mm_uuids( |
| 2158 | mm_uuids, 2, modality="image", expected_uuids=["image_123", "image_123"] |
| 2159 | ) |
| 2160 | _assert_mm_uuids(mm_uuids, 1, modality="video", expected_uuids=["video_123"]) |
| 2161 | _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=["audio_123"]) |
| 2162 | |
| 2163 | |
| 2164 | def test_parse_chat_messages_multiple_modals_with_uuids_multiple_empty_media_messages_interleave( # noqa: E501 |
| 2165 | qwen25omni_model_config_mm_interleaved, |
| 2166 | image_url, |
| 2167 | video_url, |
| 2168 | audio_url, |
| 2169 | ): |
| 2170 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2171 | [ |
| 2172 | { |
| 2173 | "role": "user", |
| 2174 | "content": [ |
| 2175 | {"type": "text", "text": "What's on this image?"}, |
| 2176 | { |
| 2177 | "type": "image_url", |
| 2178 | "image_url": None, |
| 2179 | "uuid": "image_123", |
| 2180 | }, |
| 2181 | {"type": "text", "text": "Now listen to this audio"}, |
| 2182 | { |
| 2183 | "type": "audio_url", |
| 2184 | "audio_url": None, |
| 2185 | "uuid": "audio_123", |
| 2186 | }, |
| 2187 | ], |
| 2188 | }, |
| 2189 | {"role": "assistant", "content": "Some stuff."}, |
| 2190 | { |
| 2191 | "role": "user", |
| 2192 | "content": [ |
| 2193 | {"type": "text", "text": "What's on this image?"}, |
| 2194 | { |
| 2195 | "type": "image_url", |
| 2196 | "image_url": None, |
| 2197 | "uuid": "image_123", |
| 2198 | }, |
| 2199 | {"type": "text", "text": "And what's in the video?"}, |
| 2200 | { |
| 2201 | "type": "video_url", |
| 2202 | "video_url": None, |
| 2203 | "uuid": "video_123", |
| 2204 | }, |
| 2205 | ], |
| 2206 | }, |
| 2207 | ], |
| 2208 | qwen25omni_model_config_mm_interleaved, |
| 2209 | content_format="string", |
| 2210 | ) |
| 2211 | |
| 2212 | assert conversation == [ |
| 2213 | { |
| 2214 | "role": "user", |
| 2215 | "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>" |
| 2216 | "\nNow listen to this audio\nAudio 1: <|audio_bos|><|AUDIO|><|audio_eos|>", |
| 2217 | }, |
| 2218 | {"role": "assistant", "content": "Some stuff."}, |
| 2219 | { |
| 2220 | "role": "user", |
| 2221 | "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>" |
| 2222 | "\nAnd what's in the video?\n<|vision_start|><|VIDEO|><|vision_end|>", |
| 2223 | }, |
| 2224 | ] |
| 2225 | |
| 2226 | _assert_mm_data_inputs( |
| 2227 | mm_data, |
| 2228 | {"image": 2, "video": 1, "audio": 1}, |
| 2229 | skipped_media_indices={"image": [0, 1], "video": [0], "audio": [0]}, |
| 2230 | ) |
| 2231 | _assert_mm_uuids( |
| 2232 | mm_uuids, 2, modality="image", expected_uuids=["image_123", "image_123"] |
| 2233 | ) |
| 2234 | _assert_mm_uuids(mm_uuids, 1, modality="video", expected_uuids=["video_123"]) |
| 2235 | _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=["audio_123"]) |
| 2236 | |
| 2237 | |
| 2238 | def test_parse_chat_messages_multiple_modals_with_partial_uuids_multiple_messages_interleave( # noqa: E501 |
| 2239 | qwen25omni_model_config_mm_interleaved, |
| 2240 | image_url, |
| 2241 | video_url, |
| 2242 | audio_url, |
| 2243 | ): |
| 2244 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2245 | [ |
| 2246 | { |
| 2247 | "role": "user", |
| 2248 | "content": [ |
| 2249 | {"type": "text", "text": "What's on this image?"}, |
| 2250 | { |
| 2251 | "type": "image_url", |
| 2252 | "image_url": {"url": image_url}, |
| 2253 | "uuid": "image_123", |
| 2254 | }, |
| 2255 | {"type": "text", "text": "Now listen to this audio"}, |
| 2256 | {"type": "audio_url", "audio_url": {"url": audio_url}}, |
| 2257 | ], |
| 2258 | }, |
| 2259 | {"role": "assistant", "content": "Some stuff."}, |
| 2260 | { |
| 2261 | "role": "user", |
| 2262 | "content": [ |
| 2263 | {"type": "text", "text": "What's on this image?"}, |
| 2264 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 2265 | {"type": "text", "text": "And what's in the video?"}, |
| 2266 | { |
| 2267 | "type": "video_url", |
| 2268 | "video_url": {"url": video_url}, |
| 2269 | "uuid": "video_123", |
| 2270 | }, |
| 2271 | ], |
| 2272 | }, |
| 2273 | ], |
| 2274 | qwen25omni_model_config_mm_interleaved, |
| 2275 | content_format="string", |
| 2276 | ) |
| 2277 | |
| 2278 | assert conversation == [ |
| 2279 | { |
| 2280 | "role": "user", |
| 2281 | "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>" |
| 2282 | "\nNow listen to this audio\nAudio 1: <|audio_bos|><|AUDIO|><|audio_eos|>", |
| 2283 | }, |
| 2284 | {"role": "assistant", "content": "Some stuff."}, |
| 2285 | { |
| 2286 | "role": "user", |
| 2287 | "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>" |
| 2288 | "\nAnd what's in the video?\n<|vision_start|><|VIDEO|><|vision_end|>", |
| 2289 | }, |
| 2290 | ] |
| 2291 | |
| 2292 | _assert_mm_data_inputs(mm_data, {"image": 2, "video": 1, "audio": 1}) |
| 2293 | _assert_mm_uuids(mm_uuids, 2, modality="image", expected_uuids=["image_123", None]) |
| 2294 | _assert_mm_uuids(mm_uuids, 1, modality="video", expected_uuids=["video_123"]) |
| 2295 | _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[None]) |
| 2296 | |
| 2297 | |
| 2298 | def test_parse_chat_messages_multiple_images_interleave_with_placeholders( |
| 2299 | phi3v_model_config_mm_interleaved, |
| 2300 | image_url, |
| 2301 | ): |
| 2302 | with pytest.raises( |
| 2303 | VLLMValidationError, |
| 2304 | match=r"Found more '<|image_1|>' placeholders in input prompt " |
| 2305 | "than actual multimodal data items.", |
| 2306 | ): |
| 2307 | parse_chat_messages( |
| 2308 | [ |
| 2309 | { |
| 2310 | "role": "user", |
| 2311 | "content": [ |
| 2312 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 2313 | {"type": "image_url", "image_url": {"url": image_url}}, |
| 2314 | { |
| 2315 | "type": "text", |
| 2316 | "text": "I need you to compare this image\n<|image_1|>\nand this one\n<|image_2|>\n" # noqa: E501 |
| 2317 | "Do they have differences?", |
| 2318 | }, |
| 2319 | ], |
| 2320 | } |
| 2321 | ], |
| 2322 | phi3v_model_config_mm_interleaved, |
| 2323 | content_format="string", |
| 2324 | ) |
| 2325 | |
| 2326 | |
| 2327 | def test_parse_chat_messages_include_thinking_chunk(mistral_model_config): |
| 2328 | messages = [ |
| 2329 | { |
| 2330 | "role": "system", |
| 2331 | "content": [ |
| 2332 | {"type": "text", "text": "You are a helpful assistant."}, |
| 2333 | { |
| 2334 | "type": "thinking", |
| 2335 | "closed": True, |
| 2336 | "thinking": "Only return the answer when you are confident.", |
| 2337 | }, |
| 2338 | ], |
| 2339 | }, |
| 2340 | {"role": "user", "content": "What is 2+2?"}, |
| 2341 | { |
| 2342 | "role": "assistant", |
| 2343 | "content": [ |
| 2344 | {"type": "text", "text": "Let me think about it."}, |
| 2345 | {"type": "thinking", "closed": True, "thinking": "2+2 = 4"}, |
| 2346 | { |
| 2347 | "type": "text", |
| 2348 | "text": "The answer is 4.", |
| 2349 | }, |
| 2350 | ], |
| 2351 | }, |
| 2352 | ] |
| 2353 | |
| 2354 | conversation_with_thinking, _, _ = parse_chat_messages( |
| 2355 | messages, |
| 2356 | mistral_model_config, |
| 2357 | content_format="openai", |
| 2358 | ) |
| 2359 | |
| 2360 | expected_conversation = [ |
| 2361 | { |
| 2362 | "role": "system", |
| 2363 | "content": [ |
| 2364 | {"type": "text", "text": "You are a helpful assistant."}, |
| 2365 | { |
| 2366 | "type": "text", |
| 2367 | "text": "Only return the answer when you are confident.", |
| 2368 | }, |
| 2369 | ], |
| 2370 | }, |
| 2371 | { |
| 2372 | "role": "user", |
| 2373 | "content": [{"type": "text", "text": "What is 2+2?"}], |
| 2374 | }, |
| 2375 | { |
| 2376 | "role": "assistant", |
| 2377 | "content": [ |
| 2378 | {"type": "text", "text": "Let me think about it."}, |
| 2379 | {"type": "text", "text": "2+2 = 4"}, |
| 2380 | {"type": "text", "text": "The answer is 4."}, |
| 2381 | ], |
| 2382 | }, |
| 2383 | ] |
| 2384 | |
| 2385 | assert conversation_with_thinking == expected_conversation |
| 2386 | |
| 2387 | |
| 2388 | @pytest.mark.parametrize("input_audio", [None, {}, {"data": "", "format": "wav"}]) |
| 2389 | def test_parse_chat_messages_single_empty_audio_with_uuid( |
| 2390 | qwen2_audio_model_config, |
| 2391 | input_audio, |
| 2392 | ): |
| 2393 | audio_uuid = "abcd" |
| 2394 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2395 | [ |
| 2396 | { |
| 2397 | "role": "user", |
| 2398 | "content": [ |
| 2399 | { |
| 2400 | "type": "input_audio", |
| 2401 | "input_audio": input_audio, |
| 2402 | "uuid": audio_uuid, |
| 2403 | }, |
| 2404 | {"type": "text", "text": "What does the audio say?"}, |
| 2405 | ], |
| 2406 | } |
| 2407 | ], |
| 2408 | qwen2_audio_model_config, |
| 2409 | content_format="string", |
| 2410 | ) |
| 2411 | |
| 2412 | assert conversation == [ |
| 2413 | { |
| 2414 | "role": "user", |
| 2415 | "content": "Audio 1: <|audio_bos|><|AUDIO|><|audio_eos|>\nWhat does the " |
| 2416 | "audio say?", |
| 2417 | } |
| 2418 | ] |
| 2419 | _assert_mm_data_inputs(mm_data, {"audio": 1}) |
| 2420 | _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[audio_uuid]) |
| 2421 | |
| 2422 | |
| 2423 | @pytest.mark.asyncio |
| 2424 | @pytest.mark.parametrize("input_audio", [None, {}, {"data": "", "format": "wav"}]) |
| 2425 | async def test_parse_chat_messages_single_empty_audio_with_uuid_async( |
| 2426 | qwen2_audio_model_config, |
| 2427 | input_audio, |
| 2428 | ): |
| 2429 | audio_uuid = "abcd" |
| 2430 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 2431 | [ |
| 2432 | { |
| 2433 | "role": "user", |
| 2434 | "content": [ |
| 2435 | { |
| 2436 | "type": "input_audio", |
| 2437 | "input_audio": input_audio, |
| 2438 | "uuid": audio_uuid, |
| 2439 | }, |
| 2440 | {"type": "text", "text": "What does the audio say?"}, |
| 2441 | ], |
| 2442 | } |
| 2443 | ], |
| 2444 | qwen2_audio_model_config, |
| 2445 | content_format="string", |
| 2446 | ) |
| 2447 | |
| 2448 | assert conversation == [ |
| 2449 | { |
| 2450 | "role": "user", |
| 2451 | "content": "Audio 1: <|audio_bos|><|AUDIO|><|audio_eos|>\nWhat does the " |
| 2452 | "audio say?", |
| 2453 | } |
| 2454 | ] |
| 2455 | _assert_mm_data_inputs(mm_data, {"audio": 1}) |
| 2456 | _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[audio_uuid]) |
| 2457 | |
| 2458 | |
| 2459 | def test_parse_chat_messages_image_vision_chunk( |
| 2460 | kimi_k2_5_model_config, |
| 2461 | image_url, |
| 2462 | ): |
| 2463 | messages: list[ChatCompletionMessageParam] = [ |
| 2464 | { |
| 2465 | "role": "user", |
| 2466 | "content": [ |
| 2467 | {"type": "text", "text": "Analyze this image."}, |
| 2468 | { |
| 2469 | "type": "image_url", |
| 2470 | "image_url": {"url": image_url}, |
| 2471 | }, |
| 2472 | ], |
| 2473 | } |
| 2474 | ] |
| 2475 | |
| 2476 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2477 | messages, |
| 2478 | kimi_k2_5_model_config, |
| 2479 | content_format="string", |
| 2480 | ) |
| 2481 | |
| 2482 | placeholder = "<|media_begin|>image<|media_content|><|media_pad|><|media_end|>" |
| 2483 | expected_conversation = [ |
| 2484 | { |
| 2485 | "role": "user", |
| 2486 | "content": f"{placeholder}\nAnalyze this image.", |
| 2487 | } |
| 2488 | ] |
| 2489 | |
| 2490 | assert conversation == expected_conversation |
| 2491 | _assert_mm_data_is_vision_chunk_input(mm_data, 1) |
| 2492 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None], modality="vision_chunk") |
| 2493 | |
| 2494 | |
| 2495 | def test_parse_chat_messages_video_vision_chunk( |
| 2496 | kimi_k2_5_model_config, |
| 2497 | video_url, |
| 2498 | ): |
| 2499 | messages: list[ChatCompletionMessageParam] = [ |
| 2500 | { |
| 2501 | "role": "user", |
| 2502 | "content": [ |
| 2503 | {"type": "text", "text": "Analyze this video."}, |
| 2504 | { |
| 2505 | "type": "video_url", |
| 2506 | "video_url": {"url": video_url}, |
| 2507 | }, |
| 2508 | ], |
| 2509 | } |
| 2510 | ] |
| 2511 | |
| 2512 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2513 | messages, |
| 2514 | kimi_k2_5_model_config, |
| 2515 | content_format="string", |
| 2516 | ) |
| 2517 | |
| 2518 | placeholder = "<|kimi_k25_video_placeholder|>" |
| 2519 | expected_conversation = [ |
| 2520 | { |
| 2521 | "role": "user", |
| 2522 | "content": f"{placeholder}\nAnalyze this video.", |
| 2523 | } |
| 2524 | ] |
| 2525 | |
| 2526 | assert conversation == expected_conversation |
| 2527 | _assert_mm_data_is_vision_chunk_input(mm_data, 1) |
| 2528 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None], modality="vision_chunk") |
| 2529 | |
| 2530 | |
| 2531 | def test_parse_chat_messages_image_vision_chunk_with_uuid( |
| 2532 | kimi_k2_5_model_config, |
| 2533 | image_url, |
| 2534 | ): |
| 2535 | image_uuid = "image_123" |
| 2536 | messages: list[ChatCompletionMessageParam] = [ |
| 2537 | { |
| 2538 | "role": "user", |
| 2539 | "content": [ |
| 2540 | {"type": "text", "text": "Analyze this image."}, |
| 2541 | { |
| 2542 | "type": "image_url", |
| 2543 | "image_url": {"url": image_url}, |
| 2544 | "uuid": image_uuid, |
| 2545 | }, |
| 2546 | ], |
| 2547 | } |
| 2548 | ] |
| 2549 | |
| 2550 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2551 | messages, |
| 2552 | kimi_k2_5_model_config, |
| 2553 | content_format="string", |
| 2554 | ) |
| 2555 | |
| 2556 | placeholder = "<|media_begin|>image<|media_content|><|media_pad|><|media_end|>" |
| 2557 | expected_conversation = [ |
| 2558 | { |
| 2559 | "role": "user", |
| 2560 | "content": f"{placeholder}\nAnalyze this image.", |
| 2561 | } |
| 2562 | ] |
| 2563 | |
| 2564 | assert conversation == expected_conversation |
| 2565 | _assert_mm_data_is_vision_chunk_input(mm_data, 1) |
| 2566 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[image_uuid], modality="vision_chunk") |
| 2567 | |
| 2568 | |
| 2569 | def test_parse_chat_messages_video_vision_chunk_with_uuid( |
| 2570 | kimi_k2_5_model_config, |
| 2571 | video_url, |
| 2572 | ): |
| 2573 | video_uuid = "video_456" |
| 2574 | messages: list[ChatCompletionMessageParam] = [ |
| 2575 | { |
| 2576 | "role": "user", |
| 2577 | "content": [ |
| 2578 | {"type": "text", "text": "Analyze this video."}, |
| 2579 | { |
| 2580 | "type": "video_url", |
| 2581 | "video_url": {"url": video_url}, |
| 2582 | "uuid": video_uuid, |
| 2583 | }, |
| 2584 | ], |
| 2585 | } |
| 2586 | ] |
| 2587 | |
| 2588 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2589 | messages, |
| 2590 | kimi_k2_5_model_config, |
| 2591 | content_format="string", |
| 2592 | ) |
| 2593 | |
| 2594 | placeholder = "<|kimi_k25_video_placeholder|>" |
| 2595 | expected_conversation = [ |
| 2596 | { |
| 2597 | "role": "user", |
| 2598 | "content": f"{placeholder}\nAnalyze this video.", |
| 2599 | } |
| 2600 | ] |
| 2601 | |
| 2602 | assert conversation == expected_conversation |
| 2603 | _assert_mm_data_is_vision_chunk_input(mm_data, 1) |
| 2604 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[video_uuid], modality="vision_chunk") |
| 2605 | |
| 2606 | |
| 2607 | def test_parse_chat_messages_mixed_vision_chunk( |
| 2608 | kimi_k2_5_model_config, |
| 2609 | image_url, |
| 2610 | video_url, |
| 2611 | ): |
| 2612 | messages: list[ChatCompletionMessageParam] = [ |
| 2613 | { |
| 2614 | "role": "user", |
| 2615 | "content": [ |
| 2616 | {"type": "text", "text": "Analyze this image and video."}, |
| 2617 | { |
| 2618 | "type": "image_url", |
| 2619 | "image_url": {"url": image_url}, |
| 2620 | }, |
| 2621 | { |
| 2622 | "type": "video_url", |
| 2623 | "video_url": {"url": video_url}, |
| 2624 | }, |
| 2625 | ], |
| 2626 | } |
| 2627 | ] |
| 2628 | |
| 2629 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2630 | messages, |
| 2631 | kimi_k2_5_model_config, |
| 2632 | content_format="string", |
| 2633 | ) |
| 2634 | |
| 2635 | image_placeholder = ( |
| 2636 | "<|media_begin|>image<|media_content|><|media_pad|><|media_end|>" |
| 2637 | ) |
| 2638 | video_placeholder = "<|kimi_k25_video_placeholder|>" |
| 2639 | expected_conversation = [ |
| 2640 | { |
| 2641 | "role": "user", |
| 2642 | "content": ( |
| 2643 | f"{image_placeholder}\n{video_placeholder}\n" |
| 2644 | "Analyze this image and video." |
| 2645 | ), |
| 2646 | } |
| 2647 | ] |
| 2648 | |
| 2649 | assert conversation == expected_conversation |
| 2650 | _assert_mm_data_is_vision_chunk_input(mm_data, 2) |
| 2651 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None], modality="vision_chunk") |
| 2652 | |
| 2653 | |
| 2654 | def test_parse_chat_messages_mixed_vision_chunk_with_uuid( |
| 2655 | kimi_k2_5_model_config, |
| 2656 | image_url, |
| 2657 | video_url, |
| 2658 | ): |
| 2659 | image_uuid = "image_123" |
| 2660 | video_uuid = "video_456" |
| 2661 | messages: list[ChatCompletionMessageParam] = [ |
| 2662 | { |
| 2663 | "role": "user", |
| 2664 | "content": [ |
| 2665 | {"type": "text", "text": "Analyze this image and video."}, |
| 2666 | { |
| 2667 | "type": "image_url", |
| 2668 | "image_url": {"url": image_url}, |
| 2669 | "uuid": image_uuid, |
| 2670 | }, |
| 2671 | { |
| 2672 | "type": "video_url", |
| 2673 | "video_url": {"url": video_url}, |
| 2674 | "uuid": video_uuid, |
| 2675 | }, |
| 2676 | ], |
| 2677 | } |
| 2678 | ] |
| 2679 | |
| 2680 | conversation, mm_data, mm_uuids = parse_chat_messages( |
| 2681 | messages, |
| 2682 | kimi_k2_5_model_config, |
| 2683 | content_format="string", |
| 2684 | ) |
| 2685 | |
| 2686 | image_placeholder = ( |
| 2687 | "<|media_begin|>image<|media_content|><|media_pad|><|media_end|>" |
| 2688 | ) |
| 2689 | video_placeholder = "<|kimi_k25_video_placeholder|>" |
| 2690 | expected_conversation = [ |
| 2691 | { |
| 2692 | "role": "user", |
| 2693 | "content": ( |
| 2694 | f"{image_placeholder}\n{video_placeholder}\n" |
| 2695 | "Analyze this image and video." |
| 2696 | ), |
| 2697 | } |
| 2698 | ] |
| 2699 | |
| 2700 | assert conversation == expected_conversation |
| 2701 | _assert_mm_data_is_vision_chunk_input(mm_data, 2) |
| 2702 | _assert_mm_uuids( |
| 2703 | mm_uuids, 2, expected_uuids=[image_uuid, video_uuid], modality="vision_chunk" |
| 2704 | ) |
| 2705 | |
| 2706 | |
| 2707 | @pytest.mark.asyncio |
| 2708 | async def test_parse_chat_messages_mixed_vision_chunk_async( |
| 2709 | kimi_k2_5_model_config, |
| 2710 | image_url, |
| 2711 | video_url, |
| 2712 | ): |
| 2713 | messages: list[ChatCompletionMessageParam] = [ |
| 2714 | { |
| 2715 | "role": "user", |
| 2716 | "content": [ |
| 2717 | {"type": "text", "text": "Analyze this image and video."}, |
| 2718 | { |
| 2719 | "type": "image_url", |
| 2720 | "image_url": {"url": image_url}, |
| 2721 | }, |
| 2722 | { |
| 2723 | "type": "video_url", |
| 2724 | "video_url": {"url": video_url}, |
| 2725 | }, |
| 2726 | ], |
| 2727 | } |
| 2728 | ] |
| 2729 | |
| 2730 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 2731 | messages, |
| 2732 | kimi_k2_5_model_config, |
| 2733 | content_format="string", |
| 2734 | ) |
| 2735 | |
| 2736 | image_placeholder = ( |
| 2737 | "<|media_begin|>image<|media_content|><|media_pad|><|media_end|>" |
| 2738 | ) |
| 2739 | video_placeholder = "<|kimi_k25_video_placeholder|>" |
| 2740 | expected_conversation = [ |
| 2741 | { |
| 2742 | "role": "user", |
| 2743 | "content": ( |
| 2744 | f"{image_placeholder}\n{video_placeholder}\n" |
| 2745 | "Analyze this image and video." |
| 2746 | ), |
| 2747 | } |
| 2748 | ] |
| 2749 | |
| 2750 | assert conversation == expected_conversation |
| 2751 | _assert_mm_data_is_vision_chunk_input(mm_data, 2) |
| 2752 | _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None], modality="vision_chunk") |
| 2753 | |
| 2754 | |
| 2755 | @pytest.mark.asyncio |
| 2756 | async def test_parse_chat_messages_mixed_vision_chunk_with_uuid_async( |
| 2757 | kimi_k2_5_model_config, |
| 2758 | image_url, |
| 2759 | video_url, |
| 2760 | ): |
| 2761 | image_uuid = "image_123" |
| 2762 | video_uuid = "video_456" |
| 2763 | messages: list[ChatCompletionMessageParam] = [ |
| 2764 | { |
| 2765 | "role": "user", |
| 2766 | "content": [ |
| 2767 | {"type": "text", "text": "Analyze this image and video."}, |
| 2768 | { |
| 2769 | "type": "image_url", |
| 2770 | "image_url": {"url": image_url}, |
| 2771 | "uuid": image_uuid, |
| 2772 | }, |
| 2773 | { |
| 2774 | "type": "video_url", |
| 2775 | "video_url": {"url": video_url}, |
| 2776 | "uuid": video_uuid, |
| 2777 | }, |
| 2778 | ], |
| 2779 | } |
| 2780 | ] |
| 2781 | |
| 2782 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 2783 | messages, |
| 2784 | kimi_k2_5_model_config, |
| 2785 | content_format="string", |
| 2786 | ) |
| 2787 | |
| 2788 | image_placeholder = ( |
| 2789 | "<|media_begin|>image<|media_content|><|media_pad|><|media_end|>" |
| 2790 | ) |
| 2791 | video_placeholder = "<|kimi_k25_video_placeholder|>" |
| 2792 | expected_conversation = [ |
| 2793 | { |
| 2794 | "role": "user", |
| 2795 | "content": ( |
| 2796 | f"{image_placeholder}\n{video_placeholder}\n" |
| 2797 | "Analyze this image and video." |
| 2798 | ), |
| 2799 | } |
| 2800 | ] |
| 2801 | |
| 2802 | assert conversation == expected_conversation |
| 2803 | _assert_mm_data_is_vision_chunk_input(mm_data, 2) |
| 2804 | _assert_mm_uuids( |
| 2805 | mm_uuids, 2, expected_uuids=[image_uuid, video_uuid], modality="vision_chunk" |
| 2806 | ) |
| 2807 | |
| 2808 | |
| 2809 | @pytest.mark.asyncio |
| 2810 | async def test_parse_chat_messages_image_vision_chunk_async( |
| 2811 | kimi_k2_5_model_config, |
| 2812 | image_url, |
| 2813 | ): |
| 2814 | messages: list[ChatCompletionMessageParam] = [ |
| 2815 | { |
| 2816 | "role": "user", |
| 2817 | "content": [ |
| 2818 | {"type": "text", "text": "Analyze this image."}, |
| 2819 | { |
| 2820 | "type": "image_url", |
| 2821 | "image_url": {"url": image_url}, |
| 2822 | }, |
| 2823 | ], |
| 2824 | } |
| 2825 | ] |
| 2826 | |
| 2827 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 2828 | messages, |
| 2829 | kimi_k2_5_model_config, |
| 2830 | content_format="string", |
| 2831 | ) |
| 2832 | |
| 2833 | placeholder = "<|media_begin|>image<|media_content|><|media_pad|><|media_end|>" |
| 2834 | expected_conversation = [ |
| 2835 | { |
| 2836 | "role": "user", |
| 2837 | "content": f"{placeholder}\nAnalyze this image.", |
| 2838 | } |
| 2839 | ] |
| 2840 | |
| 2841 | assert conversation == expected_conversation |
| 2842 | _assert_mm_data_is_vision_chunk_input(mm_data, 1) |
| 2843 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None], modality="vision_chunk") |
| 2844 | |
| 2845 | |
| 2846 | @pytest.mark.asyncio |
| 2847 | async def test_parse_chat_messages_video_vision_chunk_async( |
| 2848 | kimi_k2_5_model_config, |
| 2849 | video_url, |
| 2850 | ): |
| 2851 | messages: list[ChatCompletionMessageParam] = [ |
| 2852 | { |
| 2853 | "role": "user", |
| 2854 | "content": [ |
| 2855 | {"type": "text", "text": "Analyze this video."}, |
| 2856 | { |
| 2857 | "type": "video_url", |
| 2858 | "video_url": {"url": video_url}, |
| 2859 | }, |
| 2860 | ], |
| 2861 | } |
| 2862 | ] |
| 2863 | |
| 2864 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 2865 | messages, |
| 2866 | kimi_k2_5_model_config, |
| 2867 | content_format="string", |
| 2868 | ) |
| 2869 | |
| 2870 | placeholder = "<|kimi_k25_video_placeholder|>" |
| 2871 | expected_conversation = [ |
| 2872 | { |
| 2873 | "role": "user", |
| 2874 | "content": f"{placeholder}\nAnalyze this video.", |
| 2875 | } |
| 2876 | ] |
| 2877 | |
| 2878 | assert conversation == expected_conversation |
| 2879 | _assert_mm_data_is_vision_chunk_input(mm_data, 1) |
| 2880 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None], modality="vision_chunk") |
| 2881 | |
| 2882 | |
| 2883 | @pytest.mark.asyncio |
| 2884 | async def test_parse_chat_messages_image_vision_chunk_with_uuid_async( |
| 2885 | kimi_k2_5_model_config, |
| 2886 | image_url, |
| 2887 | ): |
| 2888 | image_uuid = "image_123" |
| 2889 | messages: list[ChatCompletionMessageParam] = [ |
| 2890 | { |
| 2891 | "role": "user", |
| 2892 | "content": [ |
| 2893 | {"type": "text", "text": "Analyze this image."}, |
| 2894 | { |
| 2895 | "type": "image_url", |
| 2896 | "image_url": {"url": image_url}, |
| 2897 | "uuid": image_uuid, |
| 2898 | }, |
| 2899 | ], |
| 2900 | } |
| 2901 | ] |
| 2902 | |
| 2903 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 2904 | messages, |
| 2905 | kimi_k2_5_model_config, |
| 2906 | content_format="string", |
| 2907 | ) |
| 2908 | |
| 2909 | placeholder = "<|media_begin|>image<|media_content|><|media_pad|><|media_end|>" |
| 2910 | expected_conversation = [ |
| 2911 | { |
| 2912 | "role": "user", |
| 2913 | "content": f"{placeholder}\nAnalyze this image.", |
| 2914 | } |
| 2915 | ] |
| 2916 | |
| 2917 | assert conversation == expected_conversation |
| 2918 | _assert_mm_data_is_vision_chunk_input(mm_data, 1) |
| 2919 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[image_uuid], modality="vision_chunk") |
| 2920 | |
| 2921 | |
| 2922 | @pytest.mark.asyncio |
| 2923 | async def test_parse_chat_messages_video_vision_chunk_with_uuid_async( |
| 2924 | kimi_k2_5_model_config, |
| 2925 | video_url, |
| 2926 | ): |
| 2927 | video_uuid = "video_456" |
| 2928 | messages: list[ChatCompletionMessageParam] = [ |
| 2929 | { |
| 2930 | "role": "user", |
| 2931 | "content": [ |
| 2932 | {"type": "text", "text": "Analyze this video."}, |
| 2933 | { |
| 2934 | "type": "video_url", |
| 2935 | "video_url": {"url": video_url}, |
| 2936 | "uuid": video_uuid, |
| 2937 | }, |
| 2938 | ], |
| 2939 | } |
| 2940 | ] |
| 2941 | |
| 2942 | conversation, mm_data, mm_uuids = await parse_chat_messages_async( |
| 2943 | messages, |
| 2944 | kimi_k2_5_model_config, |
| 2945 | content_format="string", |
| 2946 | ) |
| 2947 | |
| 2948 | placeholder = "<|kimi_k25_video_placeholder|>" |
| 2949 | expected_conversation = [ |
| 2950 | { |
| 2951 | "role": "user", |
| 2952 | "content": f"{placeholder}\nAnalyze this video.", |
| 2953 | } |
| 2954 | ] |
| 2955 | |
| 2956 | assert conversation == expected_conversation |
| 2957 | _assert_mm_data_is_vision_chunk_input(mm_data, 1) |
| 2958 | _assert_mm_uuids(mm_uuids, 1, expected_uuids=[video_uuid], modality="vision_chunk") |
| 2959 | |
| 2960 | |
| 2961 | def test_postprocess_messages_null_arguments_string(): |
| 2962 | """arguments="null" must not reach the chat template as Python None. |
| 2963 | |
| 2964 | json.loads("null") returns None, which causes Jinja2 templates that call |
| 2965 | tc.arguments.items() to raise 'None' has no attribute 'items'. |
| 2966 | The function should coerce it to {} instead. |
| 2967 | """ |
| 2968 | messages: list[ConversationMessage] = [ |
| 2969 | { |
| 2970 | "role": "assistant", |
| 2971 | "content": None, |
| 2972 | "tool_calls": [ |
| 2973 | { |
| 2974 | "id": "call_1", |
| 2975 | "type": "function", |
| 2976 | "function": {"name": "get_current_time", "arguments": "null"}, |
| 2977 | } |
| 2978 | ], |
| 2979 | } |
| 2980 | ] |
| 2981 | _postprocess_messages(messages) |
| 2982 | tool_calls = messages[0]["tool_calls"] |
| 2983 | assert tool_calls is not None |
| 2984 | assert tool_calls[0]["function"]["arguments"] == {} |
| 2985 | |
| 2986 | |
| 2987 | @pytest.mark.asyncio |
| 2988 | async def test_resolve_items_runs_modalities_concurrently_and_preserves_order(): |
| 2989 | """Media fetches overlap while modality and item order are preserved.""" |
| 2990 | active_fetches = 0 |
| 2991 | max_active_fetches = 0 |
| 2992 | |
| 2993 | async def _fetch(name: str, delay: float): |
| 2994 | nonlocal active_fetches, max_active_fetches |
| 2995 | active_fetches += 1 |
| 2996 | max_active_fetches = max(max_active_fetches, active_fetches) |
| 2997 | try: |
| 2998 | await asyncio.sleep(delay) |
| 2999 | return name, f"{name}-uuid" |
| 3000 | finally: |
| 3001 | active_fetches -= 1 |
| 3002 | |
| 3003 | tracker = AsyncMultiModalItemTracker(MagicMock(is_multimodal_model=True)) |
| 3004 | tracker.__dict__["mm_processor"] = MagicMock() |
| 3005 | tracker._items_by_modality["video"] = [ |
| 3006 | lambda: _fetch("video-0", 0.02), |
| 3007 | lambda: _fetch("video-1", 0.01), |
| 3008 | ] |
| 3009 | tracker._items_by_modality["image"] = [ |
| 3010 | lambda: _fetch("image-0", 0.02), |
| 3011 | lambda: _fetch("image-1", 0.01), |
| 3012 | ] |
| 3013 | tracker._items_by_modality["audio"] = [ |
| 3014 | lambda: _fetch("audio-0", 0.02), |
| 3015 | lambda: _fetch("audio-1", 0.01), |
| 3016 | ] |
| 3017 | |
| 3018 | mm_data, mm_uuids = await tracker.resolve_items() |
| 3019 | |
| 3020 | assert mm_data is not None |
| 3021 | assert mm_uuids is not None |
| 3022 | assert max_active_fetches == 6 |
| 3023 | assert list(mm_data) == ["image", "audio", "video"] |
| 3024 | assert list(mm_uuids) == ["image", "audio", "video"] |
| 3025 | assert mm_data == { |
| 3026 | "image": ["image-0", "image-1"], |
| 3027 | "audio": ["audio-0", "audio-1"], |
| 3028 | "video": ["video-0", "video-1"], |
| 3029 | } |
| 3030 | assert mm_uuids == { |
| 3031 | "image": ["image-0-uuid", "image-1-uuid"], |
| 3032 | "audio": ["audio-0-uuid", "audio-1-uuid"], |
| 3033 | "video": ["video-0-uuid", "video-1-uuid"], |
| 3034 | } |
| 3035 | |
| 3036 | |
| 3037 | @pytest.mark.asyncio |
| 3038 | async def test_resolve_items_does_not_leak_tasks_on_partial_failure(): |
| 3039 | """Regression test: one failing media fetch must not abandon the other |
| 3040 | still-in-flight fetches across modality batches. |
| 3041 | |
| 3042 | Before the fix, `resolve_items` gathered per-modality fetches with plain |
| 3043 | `asyncio.gather`, so the first exception propagated immediately while |
| 3044 | sibling fetches (real network/thread-pool work in production) kept |
| 3045 | running detached, with nothing left to await or cancel them. |
| 3046 | """ |
| 3047 | |
| 3048 | async def _fetch(should_fail: bool, delay: float): |
| 3049 | if should_fail: |
| 3050 | await asyncio.sleep(0.01) |
| 3051 | raise ValueError("simulated fetch failure") |
| 3052 | await asyncio.sleep(delay) |
| 3053 | return ("decoded", None) |
| 3054 | |
| 3055 | tracker = AsyncMultiModalItemTracker(MagicMock()) |
| 3056 | tracker._items_by_modality["image"] = [ |
| 3057 | lambda: _fetch(True, 0), |
| 3058 | lambda: _fetch(False, 0.2), |
| 3059 | lambda: _fetch(False, 0.2), |
| 3060 | ] |
| 3061 | tracker._items_by_modality["audio"] = [lambda: _fetch(False, 0.2)] |
| 3062 | tracker._items_by_modality["video"] = [lambda: _fetch(False, 0.2)] |
| 3063 | |
| 3064 | tasks_before = asyncio.all_tasks() |
| 3065 | with pytest.raises(ValueError, match="simulated fetch failure"): |
| 3066 | await tracker.resolve_items() |
| 3067 | |
| 3068 | leaked_tasks = asyncio.all_tasks() - tasks_before |
| 3069 | assert not leaked_tasks, ( |
| 3070 | f"resolve_items left {len(leaked_tasks)} task(s) running after " |
| 3071 | f"raising: {leaked_tasks}" |
| 3072 | ) |
| 3073 | |
| 3074 | |
| 3075 | def _assistant_tool_call(arguments, name="write"): |
| 3076 | return [ |
| 3077 | { |
| 3078 | "role": "assistant", |
| 3079 | "content": "", |
| 3080 | "tool_calls": [ |
| 3081 | { |
| 3082 | "id": "call_0", |
| 3083 | "type": "function", |
| 3084 | "function": {"name": name, "arguments": arguments}, |
| 3085 | } |
| 3086 | ], |
| 3087 | } |
| 3088 | ] |
| 3089 | |
| 3090 | |
| 3091 | def _assistant_tool_calls(calls): |
| 3092 | return [ |
| 3093 | { |
| 3094 | "role": "assistant", |
| 3095 | "content": "", |
| 3096 | "tool_calls": calls, |
| 3097 | } |
| 3098 | ] |
| 3099 | |
| 3100 | |
| 3101 | def test_tool_call_arguments_dict_passthrough(caplog): |
| 3102 | messages = _assistant_tool_call({"a": 1}) |
| 3103 | _postprocess_messages(messages) |
| 3104 | args = messages[0]["tool_calls"][0]["function"]["arguments"] |
| 3105 | assert args == {"a": 1} |
| 3106 | assert not caplog.records |
| 3107 | |
| 3108 | |
| 3109 | def test_tool_call_arguments_valid_object_string(caplog): |
| 3110 | messages = _assistant_tool_call('{"a": 1}') |
| 3111 | _postprocess_messages(messages) |
| 3112 | args = messages[0]["tool_calls"][0]["function"]["arguments"] |
| 3113 | assert args == {"a": 1} |
| 3114 | assert not caplog.records |
| 3115 | |
| 3116 | |
| 3117 | def test_tool_call_arguments_malformed_json_small(caplog): |
| 3118 | bad = '{"cmd": "mkdir -p /tmp/mmadtest && cat > /tmp/mmadtest' |
| 3119 | messages = _assistant_tool_call(bad, name="exec") |
| 3120 | _postprocess_messages(messages) |
| 3121 | args = messages[0]["tool_calls"][0]["function"]["arguments"] |
| 3122 | assert args == {} |
| 3123 | |
| 3124 | assert len(caplog.records) == 1 |
| 3125 | assert "exec" in caplog.records[0].message |
| 3126 | assert "coercing to an empty object" in caplog.records[0].message |
| 3127 | |
| 3128 | |
| 3129 | def test_tool_call_arguments_malformed_json_large(caplog): |
| 3130 | bad = '{"filepath": "src/main.py", "contents": "' + ("x" * 18000) |
| 3131 | messages = _assistant_tool_call(bad, name="write") |
| 3132 | _postprocess_messages(messages) |
| 3133 | args = messages[0]["tool_calls"][0]["function"]["arguments"] |
| 3134 | assert args == {} |
| 3135 | |
| 3136 | assert len(caplog.records) == 1 |
| 3137 | assert "write" in caplog.records[0].message |
| 3138 | assert "coercing to an empty object" in caplog.records[0].message |
| 3139 | |
| 3140 | |
| 3141 | @pytest.mark.parametrize("non_obj", ["[]", "42", "true", '"hello"', "null"]) |
| 3142 | def test_tool_call_arguments_valid_json_non_object(caplog, non_obj): |
| 3143 | messages = _assistant_tool_call(non_obj, name="bad_tool") |
| 3144 | _postprocess_messages(messages) |
| 3145 | args = messages[0]["tool_calls"][0]["function"]["arguments"] |
| 3146 | assert args == {} |
| 3147 | |
| 3148 | if non_obj == "null": |
| 3149 | # null parses to None, so no warning is emitted according to requirements |
| 3150 | assert not caplog.records |
| 3151 | else: |
| 3152 | assert len(caplog.records) == 1 |
| 3153 | assert "bad_tool" in caplog.records[0].message |
| 3154 | assert "not a JSON object" in caplog.records[0].message |
| 3155 | |
| 3156 | |
| 3157 | @pytest.mark.parametrize("missing", [None, ""]) |
| 3158 | def test_tool_call_arguments_missing_or_empty(caplog, missing): |
| 3159 | messages = _assistant_tool_call(missing) |
| 3160 | if missing is None: |
| 3161 | del messages[0]["tool_calls"][0]["function"]["arguments"] |
| 3162 | _postprocess_messages(messages) |
| 3163 | args = messages[0]["tool_calls"][0]["function"]["arguments"] |
| 3164 | assert args == {} |
| 3165 | assert not caplog.records |
| 3166 | |
| 3167 | |
| 3168 | def test_tool_call_arguments_multiple_independent(caplog): |
| 3169 | calls = [ |
| 3170 | { |
| 3171 | "id": "call_0", |
| 3172 | "type": "function", |
| 3173 | "function": {"name": "good", "arguments": '{"a": 1}'}, |
| 3174 | }, |
| 3175 | { |
| 3176 | "id": "call_1", |
| 3177 | "type": "function", |
| 3178 | "function": {"name": "bad", "arguments": '{"cmd": "'}, |
| 3179 | }, |
| 3180 | ] |
| 3181 | messages = _assistant_tool_calls(calls) |
| 3182 | _postprocess_messages(messages) |
| 3183 | |
| 3184 | tool_calls = messages[0]["tool_calls"] |
| 3185 | assert tool_calls[0]["function"]["arguments"] == {"a": 1} |
| 3186 | assert tool_calls[1]["function"]["arguments"] == {} |
| 3187 | |
| 3188 | assert len(caplog.records) == 1 |
| 3189 | assert "bad" in caplog.records[0].message |
| 3190 | |
| 3191 | |
| 3192 | def test_validate_chat_template_rejects_invalid_type(): |
| 3193 | """A non-str/Path chat_template is invalid user input.""" |
| 3194 | with pytest.raises( |
| 3195 | VLLMValidationError, match="not a valid chat template type" |
| 3196 | ) as exc_info: |
| 3197 | validate_chat_template(123) |
| 3198 | assert exc_info.value.parameter == "chat_template" |