test_chat_utils.py 80 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import warnings
5
from collections.abc import Mapping
6
from typing import Literal
7
8

import pytest
9
import torch
10
from mistral_common.tokens.tokenizers.base import SpecialTokenPolicy
11

12
from vllm.assets.audio import AudioAsset
13
from vllm.assets.image import ImageAsset
14
from vllm.assets.video import VideoAsset
15
from vllm.config import ModelConfig
16
17
from vllm.entrypoints.chat_utils import (
    _try_extract_ast,
18
    apply_mistral_chat_template,
19
20
21
22
23
24
25
    load_chat_template,
    parse_chat_messages,
    parse_chat_messages_futures,
    resolve_chat_template_content_format,
    resolve_chat_template_kwargs,
    resolve_hf_chat_template,
)
26
from vllm.multimodal import MultiModalDataDict, MultiModalUUIDDict
27
28
29
30
31
from vllm.multimodal.utils import (
    encode_audio_base64,
    encode_image_base64,
    encode_video_base64,
)
32
33
from vllm.tokenizers import get_tokenizer
from vllm.tokenizers.mistral import MistralTokenizer
34
from vllm.utils.serial_utils import tensor2base64
35

36
from ..models.registry import HF_EXAMPLE_MODELS
37
38
39
40
from ..utils import VLLM_PATH

EXAMPLES_DIR = VLLM_PATH / "examples"

41
PHI3V_MODEL_ID = "microsoft/Phi-3.5-vision-instruct"
42
ULTRAVOX_MODEL_ID = "fixie-ai/ultravox-v0_5-llama-3_2-1b"
43
QWEN2AUDIO_MODEL_ID = "Qwen/Qwen2-Audio-7B-Instruct"
44
QWEN2VL_MODEL_ID = "Qwen/Qwen2-VL-2B-Instruct"
45
QWEN25VL_MODEL_ID = "Qwen/Qwen2.5-VL-3B-Instruct"
46
QWEN25OMNI_MODEL_ID = "Qwen/Qwen2.5-Omni-7B"
47
QWEN3_MODEL_ID = "Qwen/Qwen3-8B"
48
LLAMA_GUARD_MODEL_ID = "meta-llama/Llama-Guard-3-1B"
49
HERMES_MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
50
MISTRAL_MODEL_ID = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
51
52


53
@pytest.fixture(scope="function")
54
def phi3v_model_config():
55
56
57
58
59
60
61
62
    return ModelConfig(
        PHI3V_MODEL_ID,
        runner="generate",
        trust_remote_code=True,
        limit_mm_per_prompt={
            "image": 2,
        },
    )
63
64


65
66
@pytest.fixture(scope="function")
def phi3v_model_config_mm_interleaved():
67
68
69
70
71
72
73
74
75
    return ModelConfig(
        PHI3V_MODEL_ID,
        runner="generate",
        trust_remote_code=True,
        interleave_mm_strings=True,
        limit_mm_per_prompt={
            "image": 2,
        },
    )
76
77


78
79
80
81
82
83
84
85
86
87
88
89
90
@pytest.fixture(scope="function")
def phi3v_model_config_image_embeds():
    return ModelConfig(
        PHI3V_MODEL_ID,
        runner="generate",
        trust_remote_code=True,
        limit_mm_per_prompt={
            "image": 2,
        },
        enable_mm_embeds=True,
    )


91
92
93
94
95
96
97
98
99
100
101
102
@pytest.fixture(scope="function")
def qwen2_audio_model_config():
    return ModelConfig(
        QWEN2AUDIO_MODEL_ID,
        runner="generate",
        trust_remote_code=True,
        limit_mm_per_prompt={
            "audio": 1,
        },
    )


103
104
105
106
107
108
109
110
111
112
113
114
115
@pytest.fixture(scope="function")
def audio_embeds_model_config():
    return ModelConfig(
        QWEN2AUDIO_MODEL_ID,
        runner="generate",
        trust_remote_code=True,
        limit_mm_per_prompt={
            "audio": 2,
        },
        enable_mm_embeds=True,
    )


116
117
@pytest.fixture(scope="function")
def qwen25omni_model_config_mm_interleaved():
118
119
120
121
122
123
124
125
126
127
    return ModelConfig(
        QWEN25OMNI_MODEL_ID,
        runner="generate",
        interleave_mm_strings=True,
        limit_mm_per_prompt={
            "image": 2,
            "audio": 1,
            "video": 1,
        },
    )
128
129


130
131
@pytest.fixture(scope="function")
def mistral_model_config():
132
133
134
135
136
137
138
    return ModelConfig(
        MISTRAL_MODEL_ID,
        runner="generate",
        limit_mm_per_prompt={
            "image": 2,
        },
    )
139
140


141
142
@pytest.fixture(scope="module")
def image_url():
143
    image = ImageAsset("cherry_blossom")
144
145
146
147
    base64 = encode_image_base64(image.pil_image)
    return f"data:image/jpeg;base64,{base64}"


148
149
@pytest.fixture(scope="module")
def video_url():
150
    video = VideoAsset("baby_reading", 1)
151
152
153
154
155
156
    base64 = encode_video_base64(video.np_ndarrays)
    return f"data:video/jpeg;base64,{base64}"


@pytest.fixture(scope="module")
def audio_url():
157
    audio = AudioAsset("mary_had_lamb")
158
159
160
161
    base64 = encode_audio_base64(*audio.audio_and_sample_rate)
    return f"data:audio/ogg;base64,{base64}"


162
def _assert_mm_data_is_image_input(
163
    mm_data: MultiModalDataDict | None,
164
    image_count: int,
165
    skipped_image_indices: list | None = None,
166
167
168
169
170
171
172
) -> None:
    assert mm_data is not None
    assert set(mm_data.keys()) == {"image"}

    image_data = mm_data.get("image")
    assert image_data is not None

173
    assert isinstance(image_data, list) and len(image_data) == image_count
174
175
176
    if skipped_image_indices is not None:
        for i in skipped_image_indices:
            assert image_data[i] is None
177
178


179
def _assert_mm_uuids(
180
    mm_uuids: MultiModalUUIDDict | None,
181
    media_count: int,
182
    expected_uuids: list[str | None],
183
184
185
186
187
188
189
190
191
    modality: str = "image",
) -> None:
    if len(expected_uuids) > 0:
        assert mm_uuids is not None
        assert modality in mm_uuids

        image_uuids = mm_uuids.get(modality)
        assert image_uuids is not None

192
        assert isinstance(image_uuids, list) and len(image_uuids) == media_count
193
194
195
196
197
198

        assert image_uuids == expected_uuids
    else:
        assert mm_uuids is None


199
200
201
202
203
ModalityType = Literal["image", "video", "audio"]
MultiModalDataCounts = Mapping[ModalityType, int]


def _assert_mm_data_inputs(
204
    mm_data: MultiModalDataDict | None,
205
    data_count: MultiModalDataCounts,
206
    skipped_media_indices: dict[str, list] | None = None,  # modality -> list[int]
207
208
209
210
211
212
213
214
215
) -> None:
    assert mm_data is not None
    assert set(data_count.keys()) == (set(mm_data.keys()))

    for modality, n in data_count.items():
        modality_data = mm_data.get(modality)
        assert modality_data is not None
        assert isinstance(modality_data, list) and len(modality_data) == n

216
        if skipped_media_indices is not None:
217
            skipped_media_indices_for_modality = skipped_media_indices.get(modality)
218
219
220
221
            assert skipped_media_indices_for_modality is not None
            for i in skipped_media_indices_for_modality:
                assert modality_data[i] is None

222

223
224
225
226
def test_parse_chat_messages_single_image(
    phi3v_model_config,
    image_url,
):
227
    conversation, mm_data, mm_uuids = parse_chat_messages(
228
229
230
231
232
233
234
235
236
        [
            {
                "role": "user",
                "content": [
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "What's in the image?"},
                ],
            }
        ],
237
        phi3v_model_config,
238
239
        content_format="string",
    )
240

241
242
243
    assert conversation == [
        {"role": "user", "content": "<|image_1|>\nWhat's in the image?"}
    ]
244
    _assert_mm_data_is_image_input(mm_data, 1)
245
246
247
248
249
250
251
252
253
    _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None])


def test_parse_chat_messages_single_image_with_uuid(
    phi3v_model_config,
    image_url,
):
    image_uuid = str(hash(image_url))
    conversation, mm_data, mm_uuids = parse_chat_messages(
254
255
256
257
258
259
260
261
262
263
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": {
                            "url": image_url,
                        },
                        "uuid": image_uuid,
264
                    },
265
266
267
268
                    {"type": "text", "text": "What's in the image?"},
                ],
            }
        ],
269
        phi3v_model_config,
270
271
272
        content_format="string",
    )

273
274
275
    assert conversation == [
        {"role": "user", "content": "<|image_1|>\nWhat's in the image?"}
    ]
276
277
278
279
    _assert_mm_data_is_image_input(mm_data, 1)
    _assert_mm_uuids(mm_uuids, 1, expected_uuids=[image_uuid])


280
281
282
283
284
285
def test_parse_chat_messages_single_empty_image_with_uuid(
    phi3v_model_config,
    image_url,
):
    image_uuid = str(hash(image_url))
    conversation, mm_data, mm_uuids = parse_chat_messages(
286
287
288
289
290
291
292
293
294
295
296
297
298
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": None,
                        "uuid": image_uuid,
                    },
                    {"type": "text", "text": "What's in the image?"},
                ],
            }
        ],
299
        phi3v_model_config,
300
301
302
        content_format="string",
    )

303
304
305
    assert conversation == [
        {"role": "user", "content": "<|image_1|>\nWhat's in the image?"}
    ]
306
307
308
309
    _assert_mm_data_is_image_input(mm_data, 1, skipped_image_indices=[0])
    _assert_mm_uuids(mm_uuids, 1, expected_uuids=[image_uuid])


310
311
312
313
314
315
def test_parse_chat_messages_single_image_with_bad_uuid_format(
    phi3v_model_config,
    image_url,
):
    image_uuid = str(hash(image_url))
    conversation, mm_data, mm_uuids = parse_chat_messages(
316
317
318
319
320
321
322
323
324
325
326
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": {
                            "url": image_url,
                            "uuid": image_uuid,
                        },
                        "bad_uuid_key": image_uuid,
327
                    },
328
329
330
331
                    {"type": "text", "text": "What's in the image?"},
                ],
            }
        ],
332
        phi3v_model_config,
333
334
335
        content_format="string",
    )

336
337
338
    assert conversation == [
        {"role": "user", "content": "<|image_1|>\nWhat's in the image?"}
    ]
339
340
341
342
343
344
345
346
347
348
349
350
    _assert_mm_data_is_image_input(mm_data, 1)
    _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None])


def test_parse_chat_messages_multiple_images_with_uuids(
    phi3v_model_config,
    image_url,
):
    image_uuid1 = "my_uuid_1"
    image_uuid2 = "my_uuid_2"

    conversation, mm_data, mm_uuids = parse_chat_messages(
351
352
353
354
355
356
357
358
359
360
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": {
                            "url": image_url,
                        },
                        "uuid": image_uuid1,
361
                    },
362
363
364
365
366
367
                    {
                        "type": "image_url",
                        "image_url": {
                            "url": image_url,
                        },
                        "uuid": image_uuid2,
368
                    },
369
370
371
372
                    {"type": "text", "text": "What's in the image?"},
                ],
            }
        ],
373
        phi3v_model_config,
374
375
376
        content_format="string",
    )

377
378
379
380
381
382
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nWhat's in the image?",
        }
    ]
383
384
385
386
    _assert_mm_data_is_image_input(mm_data, 2)
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid1, image_uuid2])


387
388
389
390
391
392
393
394
def test_parse_chat_messages_multiple_empty_images_with_uuids(
    phi3v_model_config,
    image_url,
):
    image_uuid1 = "my_uuid_1"
    image_uuid2 = "my_uuid_2"

    conversation, mm_data, mm_uuids = parse_chat_messages(
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": None,
                        "uuid": image_uuid1,
                    },
                    {
                        "type": "image_url",
                        "image_url": None,
                        "uuid": image_uuid2,
                    },
                    {"type": "text", "text": "What's in the image?"},
                ],
            }
        ],
413
        phi3v_model_config,
414
415
416
        content_format="string",
    )

417
418
419
420
421
422
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nWhat's in the image?",
        }
    ]
423
424
425
426
427
428
429
430
431
432
433
434
    _assert_mm_data_is_image_input(mm_data, 2, skipped_image_indices=[0, 1])
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid1, image_uuid2])


def test_parse_chat_messages_mixed_empty_images_with_uuids(
    phi3v_model_config,
    image_url,
):
    image_uuid1 = "my_uuid_1"
    image_uuid2 = "my_uuid_2"

    conversation, mm_data, mm_uuids = parse_chat_messages(
435
436
437
438
439
440
441
442
443
444
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": {
                            "url": image_url,
                        },
                        "uuid": image_uuid1,
445
                    },
446
447
448
449
450
451
452
453
454
                    {
                        "type": "image_url",
                        "image_url": None,
                        "uuid": image_uuid2,
                    },
                    {"type": "text", "text": "What's in the image?"},
                ],
            }
        ],
455
        phi3v_model_config,
456
457
458
        content_format="string",
    )

459
460
461
462
463
464
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nWhat's in the image?",
        }
    ]
465
466
467
468
    _assert_mm_data_is_image_input(mm_data, 2, skipped_image_indices=[1])
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid1, image_uuid2])


469
470
471
472
473
474
475
@pytest.mark.asyncio
async def test_parse_chat_messages_single_image_with_uuid_async(
    phi3v_model_config,
    image_url,
):
    image_uuid = str(hash(image_url))
    conversation, mm_future, mm_uuids = parse_chat_messages_futures(
476
477
478
479
480
481
482
483
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": {"url": image_url},
                        "uuid": image_uuid,
484
                    },
485
486
487
488
                    {"type": "text", "text": "What's in the image?"},
                ],
            }
        ],
489
        phi3v_model_config,
490
491
492
        content_format="string",
    )

493
494
495
    assert conversation == [
        {"role": "user", "content": "<|image_1|>\nWhat's in the image?"}
    ]
496
497
498
499
    _assert_mm_data_is_image_input(await mm_future, 1)
    _assert_mm_uuids(mm_uuids, 1, expected_uuids=[image_uuid])


500
501
502
503
504
505
506
@pytest.mark.asyncio
async def test_parse_chat_messages_empty_image_with_uuid_async(
    phi3v_model_config,
    image_url,
):
    image_uuid = str(hash(image_url))
    conversation, mm_future, mm_uuids = parse_chat_messages_futures(
507
508
509
510
511
512
513
514
515
516
517
518
519
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": None,
                        "uuid": image_uuid,
                    },
                    {"type": "text", "text": "What's in the image?"},
                ],
            }
        ],
520
        phi3v_model_config,
521
522
523
        content_format="string",
    )

524
525
526
527
    assert conversation == [
        {"role": "user", "content": "<|image_1|>\nWhat's in the image?"}
    ]
    _assert_mm_data_is_image_input(await mm_future, 1, skipped_image_indices=[0])
528
529
530
    _assert_mm_uuids(mm_uuids, 1, expected_uuids=[image_uuid])


531
532
533
534
535
536
537
538
539
@pytest.mark.asyncio
async def test_parse_chat_messages_multiple_images_with_uuids_async(
    phi3v_model_config,
    image_url,
):
    image_uuid1 = "my_uuid_1"
    image_uuid2 = "my_uuid_2"

    conversation, mm_future, mm_uuids = parse_chat_messages_futures(
540
541
542
543
544
545
546
547
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": {"url": image_url},
                        "uuid": image_uuid1,
548
                    },
549
550
551
552
553
554
555
556
557
                    {
                        "type": "image_pil",
                        "image_pil": ImageAsset("cherry_blossom").pil_image,
                        "uuid": image_uuid2,
                    },
                    {"type": "text", "text": "What's in these images?"},
                ],
            }
        ],
558
        phi3v_model_config,
559
560
561
        content_format="string",
    )

562
563
564
565
566
567
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?",
        }
    ]
568
569
570
571
    _assert_mm_data_is_image_input(await mm_future, 2)
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid1, image_uuid2])


572
573
574
575
576
577
578
579
580
@pytest.mark.asyncio
async def test_parse_chat_messages_multiple_empty_images_with_uuids_async(
    phi3v_model_config,
    image_url,
):
    image_uuid1 = "my_uuid_1"
    image_uuid2 = "my_uuid_2"

    conversation, mm_future, mm_uuids = parse_chat_messages_futures(
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": None,
                        "uuid": image_uuid1,
                    },
                    {
                        "type": "image_pil",
                        "image_pil": None,
                        "uuid": image_uuid2,
                    },
                    {"type": "text", "text": "What's in these images?"},
                ],
            }
        ],
599
        phi3v_model_config,
600
601
602
        content_format="string",
    )

603
604
605
606
607
608
609
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?",
        }
    ]
    _assert_mm_data_is_image_input(await mm_future, 2, skipped_image_indices=[0, 1])
610
611
612
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid1, image_uuid2])


613
614
615
616
617
618
619
620
@pytest.mark.asyncio
async def test_parse_chat_messages_multiple_images_with_partial_uuids_async(
    phi3v_model_config,
    image_url,
):
    image_uuid2 = "my_uuid_2"

    conversation, mm_future, mm_uuids = parse_chat_messages_futures(
621
622
623
624
625
626
627
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": {"url": image_url},
628
                    },
629
630
631
632
633
634
635
636
637
                    {
                        "type": "image_pil",
                        "image_pil": ImageAsset("cherry_blossom").pil_image,
                        "uuid": image_uuid2,
                    },
                    {"type": "text", "text": "What's in these images?"},
                ],
            }
        ],
638
        phi3v_model_config,
639
640
641
        content_format="string",
    )

642
643
644
645
646
647
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?",
        }
    ]
648
649
    _assert_mm_data_is_image_input(await mm_future, 2)
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, image_uuid2])
650
651


652
653
654
655
def test_parse_chat_messages_empty_system(
    mistral_model_config,
):
    # Test string format
656
    conversation, _, _ = parse_chat_messages(
657
        [
658
            {"role": "system", "content": ""},
659
660
            {
                "role": "user",
661
                "content": [{"type": "text", "text": "Who are you?"}],
662
663
            },
        ],
664
        mistral_model_config,
665
666
        content_format="string",
    )
667
    assert conversation == [
668
669
        {"role": "system", "content": ""},
        {"role": "user", "content": "Who are you?"},
670
    ]
671
672

    # Test openai format
673
    conversation, _, _ = parse_chat_messages(
674
        [
675
            {"role": "system", "content": ""},
676
677
            {
                "role": "user",
678
                "content": [{"type": "text", "text": "Who are you?"}],
679
680
            },
        ],
681
        mistral_model_config,
682
683
684
        content_format="openai",
    )
    assert conversation == [
685
686
        {"role": "system", "content": [{"type": "text", "text": ""}]},
        {"role": "user", "content": [{"type": "text", "text": "Who are you?"}]},
687
    ]
688
689


690
@pytest.mark.asyncio
691
692
693
694
async def test_parse_chat_messages_single_image_async(
    phi3v_model_config,
    image_url,
):
695
    conversation, mm_future, mm_uuids = parse_chat_messages_futures(
696
697
698
699
700
701
702
703
704
        [
            {
                "role": "user",
                "content": [
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "What's in the image?"},
                ],
            }
        ],
705
        phi3v_model_config,
706
707
        content_format="string",
    )
708

709
710
711
    assert conversation == [
        {"role": "user", "content": "<|image_1|>\nWhat's in the image?"}
    ]
712
    _assert_mm_data_is_image_input(await mm_future, 1)
713
    _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None])
714
715
716
717
718
719


def test_parse_chat_messages_multiple_images(
    phi3v_model_config,
    image_url,
):
720
    conversation, mm_data, mm_uuids = parse_chat_messages(
721
722
723
724
725
726
727
728
729
730
731
732
733
        [
            {
                "role": "user",
                "content": [
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {
                        "type": "image_pil",
                        "image_pil": ImageAsset("cherry_blossom").pil_image,
                    },
                    {"type": "text", "text": "What's in these images?"},
                ],
            }
        ],
734
        phi3v_model_config,
735
736
        content_format="string",
    )
737

738
739
740
741
742
743
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?",
        }
    ]
744
    _assert_mm_data_is_image_input(mm_data, 2)
745
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])
746
747


748
749
750
751
752
def test_parse_chat_messages_empty_pil_image_with_uuid(
    phi3v_model_config,
):
    uuid = "abcd"
    conversation, mm_data, mm_uuids = parse_chat_messages(
753
754
755
756
757
758
759
760
761
        [
            {
                "role": "user",
                "content": [
                    {"type": "image_pil", "image_pil": None, "uuid": uuid},
                    {"type": "text", "text": "What's in this image?"},
                ],
            }
        ],
762
        phi3v_model_config,
763
764
765
        content_format="string",
    )

766
767
768
769
770
771
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\nWhat's in this image?",
        }
    ]
772
773
774
775
776
    _assert_mm_data_is_image_input(mm_data, 1, skipped_image_indices=[0])
    _assert_mm_uuids(mm_uuids, 1, expected_uuids=[uuid])


def test_parse_chat_messages_empty_image_embeds_with_uuid(
777
    phi3v_model_config_image_embeds,
778
779
780
):
    uuid = "abcd"
    conversation, mm_data, mm_uuids = parse_chat_messages(
781
782
783
784
785
786
787
788
789
        [
            {
                "role": "user",
                "content": [
                    {"type": "image_embeds", "image_embeds": None, "uuid": uuid},
                    {"type": "text", "text": "What's in this image?"},
                ],
            }
        ],
790
        phi3v_model_config_image_embeds,
791
792
793
        content_format="string",
    )

794
795
796
797
798
799
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\nWhat's in this image?",
        }
    ]
800

801
802
    assert mm_data is not None
    assert "image" in mm_data
803
804
805
806
    assert isinstance(mm_data["image"], list)
    assert len(mm_data["image"]) == 1
    assert mm_data["image"][0] is None

807
808
809
    _assert_mm_uuids(mm_uuids, 1, expected_uuids=[uuid])


810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
def test_parse_chat_messages_empty_audio_embeds_with_uuid(
    audio_embeds_model_config,
):
    """Test audio_embeds with UUID (no actual embeds data)."""
    uuid = "test-audio-uuid-123"

    conversation, mm_data, mm_uuids = parse_chat_messages(
        [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "Describe this audio"},
                    {"type": "audio_embeds", "audio_embeds": None, "uuid": uuid},
                ],
            }
        ],
826
        audio_embeds_model_config,
827
828
829
830
831
832
        content_format="string",
    )

    # Should have audio in mm_data as None (UUID provided)
    assert mm_data is not None
    assert "audio" in mm_data
833
834
835
836
    assert isinstance(mm_data["audio"], list)
    assert len(mm_data["audio"]) == 1
    assert mm_data["audio"][0] is None

837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
    # UUID should be recorded
    _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[uuid])


def test_parse_chat_messages_audio_embeds_with_string(
    audio_embeds_model_config,
):
    """Test audio_embeds with base64 string embedding data."""

    import torch

    # Create a sample audio embedding tensor
    audio_embedding = torch.randn(1, 128, 768)

    # Encode it as base64
852
    base64_audio_embedding = tensor2base64(audio_embedding)
853
854
855
856
857
858
859
860
861
862
863
864
865
866

    conversation, mm_data, mm_uuids = parse_chat_messages(
        [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "Describe this audio"},
                    {
                        "type": "audio_embeds",
                        "audio_embeds": base64_audio_embedding,
                    },
                ],
            }
        ],
867
        audio_embeds_model_config,
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
        content_format="string",
    )

    # Should have audio embedding in mm_data (single tensor, not a list)
    assert mm_data is not None
    assert "audio" in mm_data
    assert isinstance(mm_data["audio"], torch.Tensor)
    assert mm_data["audio"].shape == audio_embedding.shape
    # No UUID provided
    assert mm_uuids is not None
    assert "audio" in mm_uuids
    _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[None])


@pytest.mark.asyncio
async def test_parse_chat_messages_audio_embeds_async(
    audio_embeds_model_config,
):
    """Test audio_embeds with async futures."""

    import torch

    # Create a sample audio embedding tensor
    audio_embedding = torch.randn(1, 128, 768)

    # Encode it as base64
894
    base64_audio_embedding = tensor2base64(audio_embedding)
895
896
897
898
899
900
901
902
903
904
905
906
907
908

    conversation, mm_future, mm_uuids = parse_chat_messages_futures(
        [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "Describe this audio"},
                    {
                        "type": "audio_embeds",
                        "audio_embeds": base64_audio_embedding,
                    },
                ],
            }
        ],
909
        audio_embeds_model_config,
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
        content_format="string",
    )

    # Should have audio embedding in mm_data (single tensor, not a list)
    mm_data = await mm_future
    assert mm_data is not None
    assert "audio" in mm_data
    assert isinstance(mm_data["audio"], torch.Tensor)
    assert mm_data["audio"].shape == audio_embedding.shape
    # No UUID provided
    assert mm_uuids is not None
    assert "audio" in mm_uuids
    _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[None])


925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
def test_parse_chat_messages_multiple_image_embeds(
    phi3v_model_config_image_embeds,
):
    """Test that multiple image_embeds in a single message are now supported.

    This test validates the fix for the limitation that previously only allowed
    one message with {'type': 'image_embeds'}. Now multiple image embeddings
    can be provided in a single request, similar to regular images.
    """
    # Create two sample image embedding tensors
    image_embedding_1 = torch.randn(256, 1024)
    image_embedding_2 = torch.randn(128, 1024)

    # Encode them as base64 using the convenience function
    base64_image_embedding_1 = tensor2base64(image_embedding_1)
    base64_image_embedding_2 = tensor2base64(image_embedding_2)

    conversation, mm_data, mm_uuids = parse_chat_messages(
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_embeds",
                        "image_embeds": base64_image_embedding_1,
                    },
                    {
                        "type": "image_embeds",
                        "image_embeds": base64_image_embedding_2,
                    },
                    {"type": "text", "text": "Describe these two images."},
                ],
            }
        ],
        phi3v_model_config_image_embeds,
        content_format="string",
    )

    # Verify conversation structure
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nDescribe these two images.",
        }
    ]

    # Verify mm_data contains a list of embeddings (not a single embedding)
    assert mm_data is not None
    assert "image" in mm_data
    assert isinstance(mm_data["image"], list)
    assert len(mm_data["image"]) == 2

    # Verify each embedding has the correct shape
    assert isinstance(mm_data["image"][0], torch.Tensor)
    assert mm_data["image"][0].shape == image_embedding_1.shape
    assert isinstance(mm_data["image"][1], torch.Tensor)
    assert mm_data["image"][1].shape == image_embedding_2.shape

    # Verify UUIDs (None since we didn't provide any)
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])


def test_parse_chat_messages_multiple_image_embeds_with_uuids(
    phi3v_model_config_image_embeds,
):
    """Test multiple image_embeds with UUIDs.

    This validates that UUIDs are properly tracked for multiple embeddings.
    """
    uuid1 = "image-uuid-1"
    uuid2 = "image-uuid-2"

    conversation, mm_data, mm_uuids = parse_chat_messages(
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_embeds",
                        "image_embeds": None,
                        "uuid": uuid1,
                    },
                    {
                        "type": "image_embeds",
                        "image_embeds": None,
                        "uuid": uuid2,
                    },
                    {"type": "text", "text": "Compare these images."},
                ],
            }
        ],
        phi3v_model_config_image_embeds,
        content_format="string",
    )

    # Verify conversation structure
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nCompare these images.",
        }
    ]

    # Verify mm_data contains a list with None values (UUID references)
    assert mm_data is not None
    assert "image" in mm_data
    assert isinstance(mm_data["image"], list)
    assert len(mm_data["image"]) == 2
    assert mm_data["image"][0] is None
    assert mm_data["image"][1] is None

    # Verify UUIDs are correctly tracked
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[uuid1, uuid2])


@pytest.mark.asyncio
async def test_parse_chat_messages_multiple_image_embeds_async(
    phi3v_model_config_image_embeds,
):
    """Test multiple image_embeds with async parsing.

    This validates the AsyncMultiModalItemTracker also supports multiple embeddings.
    """
    # Create two sample image embedding tensors
    image_embedding_1 = torch.randn(200, 768)
    image_embedding_2 = torch.randn(150, 768)

    # Encode them as base64 using the convenience function
    base64_image_embedding_1 = tensor2base64(image_embedding_1)
    base64_image_embedding_2 = tensor2base64(image_embedding_2)

    conversation, mm_future, mm_uuids = parse_chat_messages_futures(
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_embeds",
                        "image_embeds": base64_image_embedding_1,
                    },
                    {
                        "type": "image_embeds",
                        "image_embeds": base64_image_embedding_2,
                    },
                    {"type": "text", "text": "What do these images show?"},
                ],
            }
        ],
        phi3v_model_config_image_embeds,
        content_format="string",
    )

    # Verify conversation structure
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nWhat do these images show?",
        }
    ]

    # Await the future and verify mm_data
    mm_data = await mm_future
    assert mm_data is not None
    assert "image" in mm_data
    assert isinstance(mm_data["image"], list)
    assert len(mm_data["image"]) == 2

    # Verify each embedding has the correct shape
    assert isinstance(mm_data["image"][0], torch.Tensor)
    assert mm_data["image"][0].shape == image_embedding_1.shape
    assert isinstance(mm_data["image"][1], torch.Tensor)
    assert mm_data["image"][1].shape == image_embedding_2.shape

    # Verify UUIDs
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])


1102
1103
@pytest.mark.asyncio
async def test_parse_chat_messages_empty_image_embeds_with_uuid_async(
1104
    phi3v_model_config_image_embeds,
1105
1106
1107
):
    uuid = "abcd"
    conversation, mm_future, mm_uuids = parse_chat_messages_futures(
1108
1109
1110
1111
1112
1113
1114
1115
1116
        [
            {
                "role": "user",
                "content": [
                    {"type": "image_embeds", "image_embeds": None, "uuid": uuid},
                    {"type": "text", "text": "What's in this image?"},
                ],
            }
        ],
1117
        phi3v_model_config_image_embeds,
1118
1119
1120
        content_format="string",
    )

1121
1122
1123
1124
1125
1126
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\nWhat's in this image?",
        }
    ]
1127
1128
1129
    mm_data = await mm_future
    assert mm_data is not None
    assert "image" in mm_data
1130
1131
1132
1133
    assert isinstance(mm_data["image"], list)
    assert len(mm_data["image"]) == 1
    assert mm_data["image"][0] is None

1134
1135
1136
    _assert_mm_uuids(mm_uuids, 1, expected_uuids=[uuid])


1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
def test_parse_chat_messages_empty_dict_image_embeds(
    phi3v_model_config_image_embeds,
):
    """Test that empty dictionary for image_embeds is handled without errors."""
    conversation, mm_data, mm_uuids = parse_chat_messages(
        [
            {
                "role": "user",
                "content": [
                    {"type": "image_embeds", "image_embeds": {}},
                    {"type": "text", "text": "What's in this image?"},
                ],
            }
        ],
        phi3v_model_config_image_embeds,
        content_format="string",
    )

    # Verify conversation structure
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\nWhat's in this image?",
        }
    ]

    # Verify mm_data contains an empty dictionary of embeddings
    assert mm_data is not None
    assert "image" in mm_data
    assert isinstance(mm_data["image"], dict)
    assert len(mm_data["image"]) == 0

    # Verify UUIDs (None since we didn't provide any)
    _assert_mm_uuids(mm_uuids, 1, expected_uuids=[None])


def test_parse_chat_messages_multiple_dict_image_embeds(
    phi3v_model_config_image_embeds,
):
    """Test that multiple dictionaries for image_embeds is handled without errors."""
    # Create two sample image embedding tensors
    batch_size = 2
    image_embedding_1 = torch.randn(batch_size, 256, 1024)
    image_embedding_2 = torch.randn(batch_size, 3)

    conversation, mm_data, mm_uuids = parse_chat_messages(
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_embeds",
                        "image_embeds": {
                            "image_embedding_1": tensor2base64(p),
                            "image_embedding_2": tensor2base64(i),
                        },
                    }
                    for p, i in zip(image_embedding_1, image_embedding_2)
                ]
                + [
                    {"type": "text", "text": "Describe these two images."},
                ],
            }
        ],
        phi3v_model_config_image_embeds,
        content_format="string",
    )

    # Verify conversation structure
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nDescribe these two images.",
        }
    ]

    # Verify mm_data contains a dictionary of multi-embeddings
    assert mm_data is not None
    assert "image" in mm_data
    assert isinstance(mm_data["image"], dict)
    assert len(mm_data["image"]) == batch_size

    # Verify each embedding has the correct shape
    assert isinstance(mm_data["image"]["image_embedding_1"], torch.Tensor)
    assert mm_data["image"]["image_embedding_1"].shape == image_embedding_1.shape
    assert isinstance(mm_data["image"]["image_embedding_2"], torch.Tensor)
    assert mm_data["image"]["image_embedding_2"].shape == image_embedding_2.shape

    # Verify UUIDs (None since we didn't provide any)
    _assert_mm_uuids(mm_uuids, batch_size, expected_uuids=[None, None])


1229
@pytest.mark.asyncio
1230
1231
1232
1233
async def test_parse_chat_messages_multiple_images_async(
    phi3v_model_config,
    image_url,
):
1234
    conversation, mm_future, mm_uuids = parse_chat_messages_futures(
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
        [
            {
                "role": "user",
                "content": [
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {
                        "type": "image_pil",
                        "image_pil": ImageAsset("cherry_blossom").pil_image,
                    },
                    {"type": "text", "text": "What's in these images?"},
                ],
            }
        ],
1248
        phi3v_model_config,
1249
1250
        content_format="string",
    )
1251

1252
1253
1254
1255
1256
1257
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?",
        }
    ]
1258
    _assert_mm_data_is_image_input(await mm_future, 2)
1259
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])
1260
1261
1262
1263
1264
1265


def test_parse_chat_messages_placeholder_already_in_prompt(
    phi3v_model_config,
    image_url,
):
1266
    conversation, mm_data, mm_uuids = parse_chat_messages(
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
        [
            {
                "role": "user",
                "content": [
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {
                        "type": "text",
                        "text": "What's in <|image_1|> and how does it compare to <|image_2|>?",  # noqa: E501
                    },
                ],
            }
        ],
1280
        phi3v_model_config,
1281
1282
        content_format="string",
    )
1283
1284
1285
1286
1287
1288
    assert conversation == [
        {
            "role": "user",
            "content": "What's in <|image_1|> and how does it compare to <|image_2|>?",
        }
    ]
1289
    _assert_mm_data_is_image_input(mm_data, 2)
1290
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])
1291
1292


1293
1294
1295
1296
def test_parse_chat_messages_placeholder_one_already_in_prompt(
    phi3v_model_config,
    image_url,
):
1297
    conversation, mm_data, mm_uuids = parse_chat_messages(
1298
1299
1300
1301
1302
1303
1304
1305
        [
            {
                "role": "user",
                "content": [
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {
                        "type": "text",
1306
1307
                        "text": "What's in <|image_1|> and how does it compare to "
                        "the other one?",
1308
1309
1310
1311
                    },
                ],
            }
        ],
1312
        phi3v_model_config,
1313
1314
        content_format="string",
    )
1315

1316
1317
1318
    assert conversation == [
        {
            "role": "user",
1319
1320
            "content": "<|image_2|>\nWhat's in <|image_1|> and how does it compare to "
            "the other one?",
1321
1322
        }
    ]
1323
    _assert_mm_data_is_image_input(mm_data, 2)
1324
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])
1325
1326


1327
1328
1329
1330
def test_parse_chat_messages_multiple_images_across_messages(
    phi3v_model_config,
    image_url,
):
1331
    conversation, mm_data, mm_uuids = parse_chat_messages(
1332
1333
        [
            {
1334
                "role": "user",
1335
                "content": [
1336
1337
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "What's in this image?"},
1338
1339
                ],
            },
1340
            {"role": "assistant", "content": "Some stuff."},
1341
            {
1342
                "role": "user",
1343
                "content": [
1344
1345
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "What about this one?"},
1346
1347
1348
                ],
            },
        ],
1349
        phi3v_model_config,
1350
1351
        content_format="string",
    )
1352
1353

    assert conversation == [
1354
1355
1356
        {"role": "user", "content": "<|image_1|>\nWhat's in this image?"},
        {"role": "assistant", "content": "Some stuff."},
        {"role": "user", "content": "<|image_2|>\nWhat about this one?"},
1357
    ]
1358
    _assert_mm_data_is_image_input(mm_data, 2)
1359
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])
1360
1361


1362
def test_parse_chat_messages_multiple_images_with_uuids_across_messages(
1363
    phi3v_model_config,
1364
    image_url,
1365
):
1366
1367
    image_uuid = str(hash(image_url))
    conversation, mm_data, mm_uuids = parse_chat_messages(
1368
1369
        [
            {
1370
                "role": "user",
1371
1372
1373
                "content": [
                    {
                        "type": "image_url",
1374
                        "image_url": {"url": image_url},
1375
1376
                        "uuid": image_uuid,
                    },
1377
                    {"type": "text", "text": "What's in this image?"},
1378
                ],
1379
            },
1380
            {"role": "assistant", "content": "Some stuff."},
1381
            {
1382
                "role": "user",
1383
1384
1385
                "content": [
                    {
                        "type": "image_url",
1386
                        "image_url": {"url": image_url},
1387
1388
                        "uuid": image_uuid,
                    },
1389
                    {"type": "text", "text": "What about this one?"},
1390
1391
1392
                ],
            },
        ],
1393
        phi3v_model_config,
1394
1395
1396
1397
        content_format="string",
    )

    assert conversation == [
1398
1399
1400
        {"role": "user", "content": "<|image_1|>\nWhat's in this image?"},
        {"role": "assistant", "content": "Some stuff."},
        {"role": "user", "content": "<|image_2|>\nWhat about this one?"},
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
    ]
    _assert_mm_data_is_image_input(mm_data, 2)
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid, image_uuid])


def test_parse_chat_messages_context_text_format(
    phi3v_model_config,
):
    conversation, mm_data, mm_uuids = parse_chat_messages(
        [
            {
                "role": "user",
1413
                "content": [{"type": "text", "text": "What's in this text?"}],
1414
            },
1415
1416
            {"role": "assistant", "content": "Some stuff."},
            {"role": "user", "content": "What about this one?"},
1417
        ],
1418
        phi3v_model_config,
1419
1420
        content_format="openai",
    )
1421
1422
1423
1424

    assert conversation == [
        {
            "role": "user",
1425
            "content": [{"type": "text", "text": "What's in this text?"}],
1426
1427
1428
        },
        {
            "role": "assistant",
1429
            "content": [{"type": "text", "text": "Some stuff."}],
1430
1431
1432
        },
        {
            "role": "user",
1433
            "content": [{"type": "text", "text": "What about this one?"}],
1434
1435
        },
    ]
1436
1437
    assert mm_data is None
    assert mm_uuids is None
1438
1439


1440
1441
1442
1443
def test_parse_chat_messages_rejects_too_many_images_in_one_message(
    phi3v_model_config,
    image_url,
):
1444
1445
1446
    with warnings.catch_warnings():
        warnings.filterwarnings(
            "ignore",
1447
1448
            message="coroutine 'async_get_and_parse_image' was never awaited",
        )
1449
        with pytest.raises(ValueError, match="At most"):
1450
            parse_chat_messages(
1451
1452
1453
1454
1455
1456
1457
                [
                    {
                        "role": "user",
                        "content": [
                            {
                                "type": "image_url",
                                "image_url": {"url": image_url},
1458
                            },
1459
1460
1461
                            {
                                "type": "image_url",
                                "image_url": {"url": image_url},
1462
                            },
1463
1464
1465
                            {
                                "type": "image_url",
                                "image_url": {"url": image_url},
1466
                            },
1467
1468
1469
1470
                            {"type": "text", "text": "What's in these images?"},
                        ],
                    }
                ],
1471
                phi3v_model_config,
1472
1473
                content_format="string",
            )
1474
1475


1476
1477
1478
1479
def test_parse_chat_messages_rejects_too_many_images_across_messages(
    phi3v_model_config,
    image_url,
):
1480
1481
1482
    with warnings.catch_warnings():
        warnings.filterwarnings(
            "ignore",
1483
1484
            message="coroutine 'async_get_and_parse_image' was never awaited",
        )
1485
        with pytest.raises(ValueError, match="At most"):
1486
            parse_chat_messages(
1487
1488
                [
                    {
1489
                        "role": "user",
1490
1491
1492
                        "content": [
                            {
                                "type": "image_url",
1493
                                "image_url": {"url": image_url},
1494
                            },
1495
                            {"type": "text", "text": "What's in this image?"},
1496
1497
                        ],
                    },
1498
                    {"role": "assistant", "content": "Some stuff."},
1499
                    {
1500
                        "role": "user",
1501
1502
1503
                        "content": [
                            {
                                "type": "image_url",
1504
                                "image_url": {"url": image_url},
1505
1506
1507
                            },
                            {
                                "type": "image_url",
1508
                                "image_url": {"url": image_url},
1509
                            },
1510
                            {"type": "text", "text": "What about these two?"},
1511
1512
1513
                        ],
                    },
                ],
1514
                phi3v_model_config,
1515
1516
                content_format="string",
            )
1517
1518
1519
1520
1521
1522


def test_parse_chat_messages_multiple_images_uncommon_input(
    phi3v_model_config,
    image_url,
):
1523
    conversation, mm_data, mm_uuids = parse_chat_messages(
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
        [
            {
                "role": "user",
                "content": [
                    "What's in these images?",
                    {"image_url": image_url},
                    {"image_url": image_url},
                ],
            }
        ],
1534
        phi3v_model_config,
1535
1536
        content_format="string",
    )
1537

1538
1539
1540
1541
1542
1543
    assert conversation == [
        {
            "role": "user",
            "content": "<|image_1|>\n<|image_2|>\nWhat's in these images?",
        }
    ]
1544
    _assert_mm_data_is_image_input(mm_data, 2)
1545
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])
1546
1547


1548
1549
1550
1551
def test_parse_chat_messages_multiple_images_interleave(
    phi3v_model_config_mm_interleaved,
    image_url,
):
1552
    conversation, mm_data, mm_uuids = parse_chat_messages(
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "text",
                        "text": "I need you to compare this image",
                    },
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "and this one"},
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "Do they have differences?"},
                ],
            }
        ],
1568
        phi3v_model_config_mm_interleaved,
1569
1570
1571
        content_format="string",
    )

1572
1573
1574
1575
1576
1577
1578
    assert conversation == [
        {
            "role": "user",
            "content": "I need you to compare this image\n<|image_1|>\nand this one\n<|image_2|>\n"  # noqa: E501
            "Do they have differences?",
        }
    ]
1579
    _assert_mm_data_is_image_input(mm_data, 2)
1580
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])
1581
1582
1583
1584
1585
1586
1587


@pytest.mark.asyncio
async def test_parse_chat_messages_multiple_images_interleave_async(
    phi3v_model_config_mm_interleaved,
    image_url,
):
1588
    conversation, mm_data, mm_uuids = parse_chat_messages_futures(
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "text",
                        "text": "I need you to compare this image",
                    },
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "and this one"},
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "Do they have differences?"},
                ],
            }
        ],
1604
        phi3v_model_config_mm_interleaved,
1605
1606
1607
        content_format="string",
    )

1608
1609
1610
1611
1612
1613
1614
    assert conversation == [
        {
            "role": "user",
            "content": "I need you to compare this image\n<|image_1|>\nand this one\n<|image_2|>\n"  # noqa: E501
            "Do they have differences?",
        }
    ]
1615
    _assert_mm_data_is_image_input(await mm_data, 2)
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])


@pytest.mark.asyncio
async def test_parse_chat_messages_multiple_images_with_uuids_interleave_async(
    phi3v_model_config_mm_interleaved,
    image_url,
):
    image_uuid = str(hash(image_url))
    conversation, mm_data, mm_uuids = parse_chat_messages_futures(
1626
1627
1628
1629
1630
1631
1632
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "text",
                        "text": "I need you to compare this image",
1633
                    },
1634
1635
1636
1637
                    {
                        "type": "image_url",
                        "image_url": {"url": image_url},
                        "uuid": image_uuid,
1638
                    },
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
                    {"type": "text", "text": "and this one"},
                    {
                        "type": "image_url",
                        "image_url": {"url": image_url},
                        "uuid": image_uuid,
                    },
                    {"type": "text", "text": "Do they have differences?"},
                ],
            }
        ],
1649
        phi3v_model_config_mm_interleaved,
1650
1651
1652
        content_format="string",
    )

1653
1654
1655
1656
1657
1658
1659
    assert conversation == [
        {
            "role": "user",
            "content": "I need you to compare this image\n<|image_1|>\nand this one\n<|image_2|>\n"  # noqa: E501
            "Do they have differences?",
        }
    ]
1660
1661
    _assert_mm_data_is_image_input(await mm_data, 2)
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid, image_uuid])
1662
1663
1664
1665
1666
1667


def test_parse_chat_messages_multiple_images_multiple_messages_interleave(
    phi3v_model_config_mm_interleaved,
    image_url,
):
1668
    conversation, mm_data, mm_uuids = parse_chat_messages(
1669
1670
        [
            {
1671
                "role": "user",
1672
                "content": [
1673
1674
1675
                    {"type": "text", "text": "What's on this image?"},
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "Be accurate."},
1676
1677
                ],
            },
1678
            {"role": "assistant", "content": "Some stuff."},
1679
            {
1680
                "role": "user",
1681
                "content": [
1682
1683
                    {"type": "text", "text": "What's on this image?"},
                    {"type": "image_url", "image_url": {"url": image_url}},
1684
1685
1686
                ],
            },
        ],
1687
        phi3v_model_config_mm_interleaved,
1688
1689
1690
        content_format="string",
    )

1691
1692
1693
1694
1695
    assert conversation == [
        {
            "role": "user",
            "content": "What's on this image?\n<|image_1|>\nBe accurate.",
        },
1696
1697
        {"role": "assistant", "content": "Some stuff."},
        {"role": "user", "content": "What's on this image?\n<|image_2|>"},
1698
    ]
1699
    _assert_mm_data_is_image_input(mm_data, 2)
1700
1701
1702
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[None, None])


1703
def test_parse_chat_messages_multiple_images_with_uuids_multiple_messages_interleave(
1704
1705
1706
1707
1708
1709
1710
    phi3v_model_config_mm_interleaved,
    image_url,
):
    image_uuid = str(hash(image_url))
    conversation, mm_data, mm_uuids = parse_chat_messages(
        [
            {
1711
                "role": "user",
1712
                "content": [
1713
                    {"type": "text", "text": "What's on this image?"},
1714
1715
                    {
                        "type": "image_url",
1716
                        "image_url": {"url": image_url},
1717
1718
                        "uuid": image_uuid,
                    },
1719
                    {"type": "text", "text": "Be accurate."},
1720
1721
                ],
            },
1722
            {"role": "assistant", "content": "Some stuff."},
1723
            {
1724
                "role": "user",
1725
                "content": [
1726
                    {"type": "text", "text": "What's on this image?"},
1727
1728
                    {
                        "type": "image_url",
1729
                        "image_url": {"url": image_url},
1730
1731
1732
1733
1734
                        "uuid": image_uuid,
                    },
                ],
            },
        ],
1735
        phi3v_model_config_mm_interleaved,
1736
1737
1738
1739
1740
1741
1742
1743
        content_format="string",
    )

    assert conversation == [
        {
            "role": "user",
            "content": "What's on this image?\n<|image_1|>\nBe accurate.",
        },
1744
1745
        {"role": "assistant", "content": "Some stuff."},
        {"role": "user", "content": "What's on this image?\n<|image_2|>"},
1746
1747
1748
    ]
    _assert_mm_data_is_image_input(mm_data, 2)
    _assert_mm_uuids(mm_uuids, 2, expected_uuids=[image_uuid, image_uuid])
1749
1750
1751


def test_parse_chat_messages_multiple_modals_multiple_messages_interleave(
1752
1753
1754
1755
1756
    qwen25omni_model_config_mm_interleaved,
    image_url,
    video_url,
    audio_url,
):
1757
    conversation, mm_data, mm_uuids = parse_chat_messages(
1758
1759
        [
            {
1760
                "role": "user",
1761
                "content": [
1762
1763
1764
1765
                    {"type": "text", "text": "What's on this image?"},
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "Now listen to this audio"},
                    {"type": "audio_url", "audio_url": {"url": audio_url}},
1766
1767
                ],
            },
1768
            {"role": "assistant", "content": "Some stuff."},
1769
            {
1770
                "role": "user",
1771
                "content": [
1772
1773
1774
1775
                    {"type": "text", "text": "What's on this image?"},
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "And what's in the video?"},
                    {"type": "video_url", "video_url": {"url": video_url}},
1776
1777
1778
                ],
            },
        ],
1779
        qwen25omni_model_config_mm_interleaved,
1780
1781
1782
1783
1784
        content_format="string",
    )

    assert conversation == [
        {
1785
            "role": "user",
1786
1787
            "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>"
            "\nNow listen to this audio\nAudio 1: <|audio_bos|><|AUDIO|><|audio_eos|>",
1788
        },
1789
        {"role": "assistant", "content": "Some stuff."},
1790
        {
1791
            "role": "user",
1792
1793
            "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>"
            "\nAnd what's in the video?\n<|vision_start|><|VIDEO|><|vision_end|>",
1794
1795
        },
    ]
1796
1797

    _assert_mm_data_inputs(mm_data, {"image": 2, "video": 1, "audio": 1})
1798
    _assert_mm_uuids(mm_uuids, 2, modality="image", expected_uuids=[None, None])
1799
1800
1801
1802
    _assert_mm_uuids(mm_uuids, 1, modality="video", expected_uuids=[None])
    _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[None])


1803
def test_parse_chat_messages_multiple_modals_with_uuids_multiple_messages_interleave(
1804
1805
1806
1807
1808
1809
1810
1811
    qwen25omni_model_config_mm_interleaved,
    image_url,
    video_url,
    audio_url,
):
    conversation, mm_data, mm_uuids = parse_chat_messages(
        [
            {
1812
                "role": "user",
1813
                "content": [
1814
                    {"type": "text", "text": "What's on this image?"},
1815
1816
                    {
                        "type": "image_url",
1817
                        "image_url": {"url": image_url},
1818
1819
                        "uuid": "image_123",
                    },
1820
                    {"type": "text", "text": "Now listen to this audio"},
1821
1822
                    {
                        "type": "audio_url",
1823
                        "audio_url": {"url": audio_url},
1824
1825
1826
1827
                        "uuid": "audio_123",
                    },
                ],
            },
1828
            {"role": "assistant", "content": "Some stuff."},
1829
            {
1830
                "role": "user",
1831
                "content": [
1832
                    {"type": "text", "text": "What's on this image?"},
1833
1834
                    {
                        "type": "image_url",
1835
                        "image_url": {"url": image_url},
1836
1837
                        "uuid": "image_123",
                    },
1838
                    {"type": "text", "text": "And what's in the video?"},
1839
1840
                    {
                        "type": "video_url",
1841
                        "video_url": {"url": video_url},
1842
1843
1844
1845
1846
                        "uuid": "video_123",
                    },
                ],
            },
        ],
1847
        qwen25omni_model_config_mm_interleaved,
1848
1849
1850
1851
1852
        content_format="string",
    )

    assert conversation == [
        {
1853
            "role": "user",
1854
1855
            "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>"
            "\nNow listen to this audio\nAudio 1: <|audio_bos|><|AUDIO|><|audio_eos|>",
1856
        },
1857
        {"role": "assistant", "content": "Some stuff."},
1858
        {
1859
            "role": "user",
1860
1861
            "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>"
            "\nAnd what's in the video?\n<|vision_start|><|VIDEO|><|vision_end|>",
1862
1863
1864
1865
        },
    ]

    _assert_mm_data_inputs(mm_data, {"image": 2, "video": 1, "audio": 1})
1866
1867
1868
1869
1870
    _assert_mm_uuids(
        mm_uuids, 2, modality="image", expected_uuids=["image_123", "image_123"]
    )
    _assert_mm_uuids(mm_uuids, 1, modality="video", expected_uuids=["video_123"])
    _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=["audio_123"])
1871
1872


1873
1874
1875
1876
1877
1878
1879
1880
1881
def test_parse_chat_messages_multiple_modals_with_uuids_multiple_empty_media_messages_interleave(  # noqa: E501
    qwen25omni_model_config_mm_interleaved,
    image_url,
    video_url,
    audio_url,
):
    conversation, mm_data, mm_uuids = parse_chat_messages(
        [
            {
1882
                "role": "user",
1883
                "content": [
1884
                    {"type": "text", "text": "What's on this image?"},
1885
1886
1887
1888
1889
                    {
                        "type": "image_url",
                        "image_url": None,
                        "uuid": "image_123",
                    },
1890
                    {"type": "text", "text": "Now listen to this audio"},
1891
1892
1893
1894
1895
1896
1897
                    {
                        "type": "audio_url",
                        "audio_url": None,
                        "uuid": "audio_123",
                    },
                ],
            },
1898
            {"role": "assistant", "content": "Some stuff."},
1899
            {
1900
                "role": "user",
1901
                "content": [
1902
                    {"type": "text", "text": "What's on this image?"},
1903
1904
1905
1906
1907
                    {
                        "type": "image_url",
                        "image_url": None,
                        "uuid": "image_123",
                    },
1908
                    {"type": "text", "text": "And what's in the video?"},
1909
1910
1911
1912
1913
1914
1915
1916
                    {
                        "type": "video_url",
                        "video_url": None,
                        "uuid": "video_123",
                    },
                ],
            },
        ],
1917
        qwen25omni_model_config_mm_interleaved,
1918
1919
1920
1921
1922
        content_format="string",
    )

    assert conversation == [
        {
1923
            "role": "user",
1924
1925
            "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>"
            "\nNow listen to this audio\nAudio 1: <|audio_bos|><|AUDIO|><|audio_eos|>",
1926
        },
1927
        {"role": "assistant", "content": "Some stuff."},
1928
        {
1929
            "role": "user",
1930
1931
            "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>"
            "\nAnd what's in the video?\n<|vision_start|><|VIDEO|><|vision_end|>",
1932
1933
1934
        },
    ]

1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
    _assert_mm_data_inputs(
        mm_data,
        {"image": 2, "video": 1, "audio": 1},
        skipped_media_indices={"image": [0, 1], "video": [0], "audio": [0]},
    )
    _assert_mm_uuids(
        mm_uuids, 2, modality="image", expected_uuids=["image_123", "image_123"]
    )
    _assert_mm_uuids(mm_uuids, 1, modality="video", expected_uuids=["video_123"])
    _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=["audio_123"])
1945
1946


1947
1948
1949
1950
1951
1952
1953
1954
1955
def test_parse_chat_messages_multiple_modals_with_partial_uuids_multiple_messages_interleave(  # noqa: E501
    qwen25omni_model_config_mm_interleaved,
    image_url,
    video_url,
    audio_url,
):
    conversation, mm_data, mm_uuids = parse_chat_messages(
        [
            {
1956
                "role": "user",
1957
                "content": [
1958
                    {"type": "text", "text": "What's on this image?"},
1959
1960
                    {
                        "type": "image_url",
1961
                        "image_url": {"url": image_url},
1962
1963
                        "uuid": "image_123",
                    },
1964
1965
                    {"type": "text", "text": "Now listen to this audio"},
                    {"type": "audio_url", "audio_url": {"url": audio_url}},
1966
1967
                ],
            },
1968
            {"role": "assistant", "content": "Some stuff."},
1969
            {
1970
                "role": "user",
1971
                "content": [
1972
1973
1974
                    {"type": "text", "text": "What's on this image?"},
                    {"type": "image_url", "image_url": {"url": image_url}},
                    {"type": "text", "text": "And what's in the video?"},
1975
1976
                    {
                        "type": "video_url",
1977
                        "video_url": {"url": video_url},
1978
1979
1980
1981
1982
                        "uuid": "video_123",
                    },
                ],
            },
        ],
1983
        qwen25omni_model_config_mm_interleaved,
1984
1985
1986
1987
1988
        content_format="string",
    )

    assert conversation == [
        {
1989
            "role": "user",
1990
1991
            "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>"
            "\nNow listen to this audio\nAudio 1: <|audio_bos|><|AUDIO|><|audio_eos|>",
1992
        },
1993
        {"role": "assistant", "content": "Some stuff."},
1994
        {
1995
            "role": "user",
1996
1997
            "content": "What's on this image?\n<|vision_start|><|IMAGE|><|vision_end|>"
            "\nAnd what's in the video?\n<|vision_start|><|VIDEO|><|vision_end|>",
1998
1999
2000
2001
        },
    ]

    _assert_mm_data_inputs(mm_data, {"image": 2, "video": 1, "audio": 1})
2002
2003
    _assert_mm_uuids(mm_uuids, 2, modality="image", expected_uuids=["image_123", None])
    _assert_mm_uuids(mm_uuids, 1, modality="video", expected_uuids=["video_123"])
2004
    _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[None])
2005
2006
2007
2008
2009
2010
2011


def test_parse_chat_messages_multiple_images_interleave_with_placeholders(
    phi3v_model_config_mm_interleaved,
    image_url,
):
    with pytest.raises(
2012
2013
2014
        ValueError,
        match=r"Found more '<|image_1|>' placeholders in input prompt "
        "than actual multimodal data items.",
2015
    ):
2016
        parse_chat_messages(
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
            [
                {
                    "role": "user",
                    "content": [
                        {"type": "image_url", "image_url": {"url": image_url}},
                        {"type": "image_url", "image_url": {"url": image_url}},
                        {
                            "type": "text",
                            "text": "I need you to compare this image\n<|image_1|>\nand this one\n<|image_2|>\n"  # noqa: E501
                            "Do they have differences?",
                        },
                    ],
                }
            ],
2031
            phi3v_model_config_mm_interleaved,
2032
2033
2034
2035
            content_format="string",
        )


2036
2037
2038
2039
2040
@pytest.mark.parametrize(
    "model",
    [
        QWEN2VL_MODEL_ID,  # tokenizer.chat_template is of type str
        HERMES_MODEL_ID,  # tokenizer.chat_template is of type dict
2041
2042
    ],
)
2043
2044
2045
@pytest.mark.parametrize("use_tools", [True, False])
def test_resolve_hf_chat_template(sample_json_schema, model, use_tools):
    """checks that chat_template is a dict type for HF models."""
2046
2047
2048
    model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
    model_info.check_available_online(on_fail="skip")

2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
    model_config = ModelConfig(
        model,
        tokenizer=model_info.tokenizer or model,
        tokenizer_mode=model_info.tokenizer_mode,
        revision=model_info.revision,
        trust_remote_code=model_info.trust_remote_code,
        hf_overrides=model_info.hf_overrides,
        skip_tokenizer_init=model_info.require_embed_inputs,
        enable_prompt_embeds=model_info.require_embed_inputs,
        enable_mm_embeds=model_info.require_embed_inputs,
        enforce_eager=model_info.enforce_eager,
        dtype=model_info.dtype,
    )

    # Build the tokenizer
2064
    tokenizer = get_tokenizer(
2065
2066
        model,
        trust_remote_code=model_config.trust_remote_code,
2067
2068
    )

2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
    tools = (
        [
            {
                "type": "function",
                "function": {
                    "name": "dummy_function_name",
                    "description": "This is a dummy function",
                    "parameters": sample_json_schema,
                },
            }
        ]
        if use_tools
        else None
    )
2083
2084

    # Test detecting the tokenizer's chat_template
2085
    chat_template = resolve_hf_chat_template(
2086
2087
2088
        tokenizer,
        chat_template=None,
        tools=tools,
2089
        model_config=model_config,
2090
2091
2092
2093
    )
    assert isinstance(chat_template, str)


2094
2095
2096
2097
2098
2099
@pytest.mark.parametrize(
    "model, expected_kwargs",
    [
        (
            QWEN2VL_MODEL_ID,
            {
2100
2101
2102
2103
                "add_vision_id",
                "add_generation_prompt",
                "continue_final_message",
                "tools",
2104
2105
2106
2107
2108
            },
        ),
        (
            QWEN3_MODEL_ID,
            {
2109
2110
2111
2112
                "enable_thinking",
                "add_generation_prompt",
                "continue_final_message",
                "tools",
2113
2114
2115
2116
            },
        ),
    ],
)
2117
def test_resolve_hf_chat_template_kwargs(sample_json_schema, model, expected_kwargs):
2118
2119
2120
2121
    """checks that chat_template is a dict type for HF models."""
    model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
    model_info.check_available_online(on_fail="skip")

2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
    tools = [
        {
            "type": "function",
            "function": {
                "name": "dummy_function_name",
                "description": "This is a dummy function",
                "parameters": sample_json_schema,
            },
        }
    ]
2132
2133
2134
2135
2136
2137
2138

    chat_template_kwargs = {
        # both unused
        "unsed_kwargs_1": 123,
        "unsed_kwargs_2": "abc",
        # should not appear
        "chat_template": "{% Hello world! %}",
2139
        "tokenize": True,
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
        # used by tokenizer
        "continue_final_message": True,
        "tools": tools,
        # both used by Qwen2-VL and Qwen3
        "add_generation_prompt": True,
        # only used by Qwen2-VL
        "add_vision_id": True,
        # only used by Qwen3
        "enable_thinking": True,
    }

2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
    model_config = ModelConfig(
        model,
        tokenizer=model_info.tokenizer or model,
        tokenizer_mode=model_info.tokenizer_mode,
        revision=model_info.revision,
        trust_remote_code=model_info.trust_remote_code,
        hf_overrides=model_info.hf_overrides,
        skip_tokenizer_init=model_info.require_embed_inputs,
        enable_prompt_embeds=model_info.require_embed_inputs,
        enable_mm_embeds=model_info.require_embed_inputs,
        enforce_eager=model_info.enforce_eager,
        dtype=model_info.dtype,
    )

    # Build the tokenizer
2166
    tokenizer = get_tokenizer(
2167
2168
        model,
        trust_remote_code=model_config.trust_remote_code,
2169
2170
2171
2172
2173
2174
2175
    )

    # Test detecting the tokenizer's chat_template
    chat_template = resolve_hf_chat_template(
        tokenizer,
        chat_template=None,
        tools=tools,
2176
        model_config=model_config,
2177
    )
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
    with pytest.raises(
        ValueError, match="Found unexpected chat template kwargs from request"
    ):
        # should raise error if `chat_template_kwargs` contains
        # `chat_template` or `tokenize`
        resolve_chat_template_kwargs(
            tokenizer,
            chat_template=chat_template,
            chat_template_kwargs=chat_template_kwargs,
        )
2188
2189
2190
2191
    resolved_chat_template_kwargs = resolve_chat_template_kwargs(
        tokenizer,
        chat_template=chat_template,
        chat_template_kwargs=chat_template_kwargs,
2192
        raise_on_unexpected=False,
2193
2194
2195
    )
    assert set(resolved_chat_template_kwargs.keys()) == expected_kwargs

2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
    # Additional test: Verify HF base parameters work with **kwargs tokenizers
    # This validates the fix for tokenizers like Kimi K2 that use **kwargs
    # to receive standard HuggingFace parameters instead of declaring them explicitly
    from vllm.entrypoints.chat_utils import _get_hf_base_chat_template_params

    hf_base_params = _get_hf_base_chat_template_params()
    # Verify common HF parameters are in the base class
    assert {"add_generation_prompt", "tools", "continue_final_message"}.issubset(
        hf_base_params
    ), f"Expected HF base params not found in {hf_base_params}"

    # Test with a mock tokenizer that uses **kwargs (like Kimi K2)
    class MockTokenizerWithKwargs:
        def apply_chat_template(self, conversation, **kwargs):
            return "mocked_output"

    mock_tokenizer = MockTokenizerWithKwargs()
    mock_kwargs = {
        "add_generation_prompt": True,
        "tools": tools,
        "continue_final_message": False,
        "unknown_param": "should_be_filtered",
    }
    resolved_mock = resolve_chat_template_kwargs(
        mock_tokenizer, chat_template, mock_kwargs, raise_on_unexpected=False
    )
    # HF base params should pass through even with **kwargs tokenizer
    assert "add_generation_prompt" in resolved_mock
    assert "tools" in resolved_mock
    assert "continue_final_message" in resolved_mock
    # Unknown params should be filtered out
    assert "unknown_param" not in resolved_mock

2229

2230
2231
# NOTE: Qwen2-Audio default chat template is specially defined inside
# processor class instead of using `tokenizer_config.json`
2232
2233
@pytest.mark.parametrize(
    ("model", "expected_format"),
2234
2235
2236
2237
2238
2239
2240
2241
    [
        (PHI3V_MODEL_ID, "string"),
        (QWEN2VL_MODEL_ID, "openai"),
        (QWEN25VL_MODEL_ID, "openai"),
        (ULTRAVOX_MODEL_ID, "string"),
        (QWEN2AUDIO_MODEL_ID, "openai"),
        (LLAMA_GUARD_MODEL_ID, "openai"),
    ],
2242
2243
)
def test_resolve_content_format_hf_defined(model, expected_format):
2244
2245
2246
    model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
    model_info.check_available_online(on_fail="skip")

2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
    model_config = ModelConfig(
        model,
        tokenizer=model_info.tokenizer or model,
        tokenizer_mode=model_info.tokenizer_mode,
        revision=model_info.revision,
        trust_remote_code=model_info.trust_remote_code,
        hf_overrides=model_info.hf_overrides,
        skip_tokenizer_init=model_info.require_embed_inputs,
        enable_prompt_embeds=model_info.require_embed_inputs,
        enable_mm_embeds=model_info.require_embed_inputs,
        enforce_eager=model_info.enforce_eager,
        dtype=model_info.dtype,
    )
2260

2261
    tokenizer = get_tokenizer(
2262
2263
        model,
        trust_remote_code=model_config.trust_remote_code,
2264
2265
    )

2266
    # Test detecting the tokenizer's chat_template
2267
    chat_template = resolve_hf_chat_template(
2268
2269
2270
        tokenizer,
        chat_template=None,
        tools=None,
2271
        model_config=model_config,
2272
    )
2273
2274
2275
2276
2277
2278
2279
2280
    assert isinstance(chat_template, str)

    print("[TEXT]")
    print(chat_template)
    print("[AST]")
    print(_try_extract_ast(chat_template))

    resolved_format = resolve_chat_template_content_format(
2281
2282
2283
2284
        None,  # Test detecting the tokenizer's chat_template
        None,
        "auto",
        tokenizer,
2285
        model_config=model_config,
2286
2287
2288
2289
2290
2291
2292
    )

    assert resolved_format == expected_format


@pytest.mark.parametrize(
    ("model", "expected_format"),
2293
2294
2295
2296
2297
2298
2299
2300
2301
    [
        ("Salesforce/blip2-opt-2.7b", "string"),
        ("facebook/chameleon-7b", "string"),
        ("deepseek-ai/deepseek-vl2-tiny", "string"),
        ("adept/fuyu-8b", "string"),
        ("google/paligemma-3b-mix-224", "string"),
        ("Qwen/Qwen-VL", "string"),
        ("Qwen/Qwen-VL-Chat", "string"),
    ],
2302
2303
2304
2305
2306
)
def test_resolve_content_format_fallbacks(model, expected_format):
    model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
    model_info.check_available_online(on_fail="skip")

2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
    model_config = ModelConfig(
        model,
        tokenizer=model_info.tokenizer or model,
        tokenizer_mode=model_info.tokenizer_mode,
        revision=model_info.revision,
        trust_remote_code=model_info.trust_remote_code,
        hf_overrides=model_info.hf_overrides,
        skip_tokenizer_init=model_info.require_embed_inputs,
        enable_prompt_embeds=model_info.require_embed_inputs,
        enable_mm_embeds=model_info.require_embed_inputs,
        enforce_eager=model_info.enforce_eager,
        dtype=model_info.dtype,
    )
2320

2321
    tokenizer = get_tokenizer(
2322
2323
        model_config.tokenizer,
        trust_remote_code=model_config.trust_remote_code,
2324
2325
2326
2327
2328
2329
2330
    )

    # Test detecting the tokenizer's chat_template
    chat_template = resolve_hf_chat_template(
        tokenizer,
        chat_template=None,
        tools=None,
2331
        model_config=model_config,
2332
2333
2334
2335
2336
2337
2338
2339
2340
    )
    assert isinstance(chat_template, str)

    print("[TEXT]")
    print(chat_template)
    print("[AST]")
    print(_try_extract_ast(chat_template))

    resolved_format = resolve_chat_template_content_format(
2341
        None,  # Test detecting the tokenizer's chat_template
2342
        None,
2343
2344
        "auto",
        tokenizer,
2345
        model_config=model_config,
2346
2347
2348
2349
2350
2351
2352
    )

    assert resolved_format == expected_format


@pytest.mark.parametrize(
    ("template_path", "expected_format"),
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
    [
        ("template_alpaca.jinja", "string"),
        ("template_baichuan.jinja", "string"),
        ("template_chatglm.jinja", "string"),
        ("template_chatglm2.jinja", "string"),
        ("template_chatml.jinja", "string"),
        ("template_dse_qwen2_vl.jinja", "openai"),
        ("template_falcon_180b.jinja", "string"),
        ("template_falcon.jinja", "string"),
        ("template_inkbot.jinja", "string"),
        ("template_teleflm.jinja", "string"),
        ("template_vlm2vec_phi3v.jinja", "openai"),
        ("template_vlm2vec_qwen2vl.jinja", "openai"),
        ("tool_chat_template_granite_20b_fc.jinja", "string"),
        ("tool_chat_template_hermes.jinja", "string"),
        ("tool_chat_template_internlm2_tool.jinja", "string"),
        ("tool_chat_template_llama3.1_json.jinja", "openai"),
        ("tool_chat_template_llama3.2_json.jinja", "openai"),
        ("tool_chat_template_mistral_parallel.jinja", "string"),
        ("tool_chat_template_mistral.jinja", "string"),
    ],
2374
2375
)
def test_resolve_content_format_examples(template_path, expected_format):
2376
2377
2378
2379
2380
    model_config = ModelConfig(
        PHI3V_MODEL_ID,  # Dummy
        tokenizer=PHI3V_MODEL_ID,  # Dummy
        trust_remote_code=True,
    )
2381

2382
    dummy_tokenizer = get_tokenizer(
2383
2384
        PHI3V_MODEL_ID,  # Dummy
        trust_remote_code=model_config.trust_remote_code,
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
    )
    dummy_tokenizer.chat_template = None

    chat_template = load_chat_template(EXAMPLES_DIR / template_path)
    assert isinstance(chat_template, str)

    print("[TEXT]")
    print(chat_template)
    print("[AST]")
    print(_try_extract_ast(chat_template))

    resolved_format = resolve_chat_template_content_format(
        chat_template,
2398
        None,
2399
2400
        "auto",
        dummy_tokenizer,
2401
        model_config=model_config,
2402
2403
2404
    )

    assert resolved_format == expected_format
Julien Denize's avatar
Julien Denize committed
2405
2406


2407
def test_parse_chat_messages_include_thinking_chunk(mistral_model_config):
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
    messages = [
        {
            "role": "system",
            "content": [
                {"type": "text", "text": "You are a helpful assistant."},
                {
                    "type": "thinking",
                    "closed": True,
                    "thinking": "Only return the answer when you are confident.",
                },
            ],
        },
        {"role": "user", "content": "What is 2+2?"},
        {
            "role": "assistant",
            "content": [
                {"type": "text", "text": "Let me think about it."},
                {"type": "thinking", "closed": True, "thinking": "2+2 = 4"},
                {
                    "type": "text",
                    "text": "The answer is 4.",
                },
            ],
        },
    ]
Julien Denize's avatar
Julien Denize committed
2433

2434
    conversation_with_thinking, _, _ = parse_chat_messages(
Julien Denize's avatar
Julien Denize committed
2435
        messages,
2436
        mistral_model_config,
Julien Denize's avatar
Julien Denize committed
2437
2438
2439
        content_format="openai",
    )

2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
    expected_conversation = [
        {
            "role": "system",
            "content": [
                {"type": "text", "text": "You are a helpful assistant."},
                {
                    "type": "text",
                    "text": "Only return the answer when you are confident.",
                },
            ],
        },
        {
            "role": "user",
            "content": [{"type": "text", "text": "What is 2+2?"}],
        },
        {
            "role": "assistant",
            "content": [
                {"type": "text", "text": "Let me think about it."},
                {"type": "text", "text": "2+2 = 4"},
                {"type": "text", "text": "The answer is 4."},
            ],
        },
    ]
Julien Denize's avatar
Julien Denize committed
2464
2465
2466
2467
2468

    assert conversation_with_thinking == expected_conversation


def test_apply_mistral_chat_template_thinking_chunk():
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
    messages = [
        {
            "role": "system",
            "content": [
                {"type": "text", "text": "You are a helpful assistant."},
                {
                    "type": "thinking",
                    "closed": True,
                    "thinking": "Only return the answer when you are confident.",
                },
            ],
        },
        {"role": "user", "content": "What is 2+2?"},
        {
            "role": "assistant",
            "content": [
                {"type": "text", "text": "Let me think about it."},
                {"type": "thinking", "closed": True, "thinking": "2+2 = 4"},
                {
                    "type": "text",
                    "text": "The answer is 4.",
                },
            ],
        },
        {"role": "user", "content": "Thanks, what is 3+3?"},
    ]
Julien Denize's avatar
Julien Denize committed
2495
    mistral_tokenizer = MistralTokenizer.from_pretrained(
2496
        "mistralai/Magistral-Small-2509"
2497
    )
Julien Denize's avatar
Julien Denize committed
2498

2499
2500
2501
    tokens_ids = apply_mistral_chat_template(
        mistral_tokenizer, messages, chat_template=None, tools=None
    )
Julien Denize's avatar
Julien Denize committed
2502
2503

    string_tokens = mistral_tokenizer.mistral.decode(
2504
2505
        tokens_ids, special_token_policy=SpecialTokenPolicy.KEEP
    )
Julien Denize's avatar
Julien Denize committed
2506
2507
2508
2509
2510
2511

    expected_tokens = (
        r"<s>[SYSTEM_PROMPT]You are a helpful assistant.[THINK]Only return the"
        r" answer when you are confident.[/THINK][/SYSTEM_PROMPT]"
        r"[INST]What is 2+2?[/INST]"
        r"Let me think about it.[THINK]2+2 = 4[/THINK]The answer is 4.</s>"
2512
2513
        r"[INST]Thanks, what is 3+3?[/INST]"
    )
Julien Denize's avatar
Julien Denize committed
2514
2515

    assert string_tokens == expected_tokens
2516
2517
2518
2519
2520
2521
2522


def test_parse_chat_messages_single_empty_audio_with_uuid(
    qwen2_audio_model_config,
):
    audio_uuid = "abcd"
    conversation, mm_data, mm_uuids = parse_chat_messages(
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "input_audio",
                        "input_audio": {},
                        "uuid": audio_uuid,
                    },
                    {"type": "text", "text": "What does the audio say?"},
                ],
            }
        ],
2536
        qwen2_audio_model_config,
2537
2538
2539
        content_format="string",
    )

2540
2541
2542
    assert conversation == [
        {
            "role": "user",
2543
2544
            "content": "Audio 1: <|audio_bos|><|AUDIO|><|audio_eos|>\nWhat does the "
            "audio say?",
2545
2546
        }
    ]
2547
    _assert_mm_data_inputs(mm_data, {"audio": 1})
2548
    _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[audio_uuid])
2549
2550
2551
2552
2553
2554
2555
2556


@pytest.mark.asyncio
async def test_parse_chat_messages_single_empty_audio_with_uuid_async(
    qwen2_audio_model_config,
):
    audio_uuid = "abcd"
    conversation, mm_future, mm_uuids = parse_chat_messages_futures(
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
        [
            {
                "role": "user",
                "content": [
                    {
                        "type": "input_audio",
                        "input_audio": {},
                        "uuid": audio_uuid,
                    },
                    {"type": "text", "text": "What does the audio say?"},
                ],
            }
        ],
2570
        qwen2_audio_model_config,
2571
2572
2573
        content_format="string",
    )

2574
2575
2576
    assert conversation == [
        {
            "role": "user",
2577
2578
            "content": "Audio 1: <|audio_bos|><|AUDIO|><|audio_eos|>\nWhat does the "
            "audio say?",
2579
2580
        }
    ]
2581
    _assert_mm_data_inputs(await mm_future, {"audio": 1})
2582
    _assert_mm_uuids(mm_uuids, 1, modality="audio", expected_uuids=[audio_uuid])