test_processing.py 32.3 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
from contextlib import nullcontext
5
from typing import Optional, cast
6

7
import numpy as np
8
import pytest
9
10
11

from vllm.config import ModelConfig
from vllm.multimodal import MULTIMODAL_REGISTRY
12

13
14
# yapf conflicts with isort for this block
# yapf: disable
15
16
17
18
19
20
21
22
23
24
25
26
27
from vllm.multimodal.processing import (
    InputProcessingContext,
    PlaceholderFeaturesInfo,
    PromptIndexTargets,
    PromptInsertion,
    PromptReplacement,
    apply_text_matches,
    apply_token_matches,
    find_mm_placeholders,
    iter_token_matches,
    replace_token_matches,
)

28
# yapf: enable
29
from vllm.multimodal.profiling import MultiModalProfiler
30
from vllm.transformers_utils.tokenizer import AnyTokenizer
31

32
33
from .utils import random_image

34
35
pytestmark = pytest.mark.cpu_test

36
37
38
39
40

# yapf: disable
@pytest.mark.parametrize(
    ("token_ids", "match_ids", "expected"),
    [
41
        ([], [], []),
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
        ([], [32000], []),
        (
            [32000, 32000, 32000],
            [32000],
            [
                { "start_idx": 0, "end_idx": 1 },
                { "start_idx": 1, "end_idx": 2 },
                { "start_idx": 2, "end_idx": 3 },
            ],
        ),
        (
            [32000, 32000, 32000],
            [32000, 32000],
            [{ "start_idx": 0, "end_idx": 2 }],
        ),
        (
            [32000, 32000, 32000],
            [32000, 32000, 32000],
            [{ "start_idx": 0, "end_idx": 3 }],
        ),
        (
            [9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918],
            [28747, 32000],
            [
                { "start_idx": 1, "end_idx": 3 },
                { "start_idx": 6, "end_idx": 8 },
            ],
        ),
        (
            [9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918],
            [28747, 32000, 32000, 32000],
            [
                { "start_idx": 1, "end_idx": 5 },
            ],
        ),
        (
            [9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918],
            [28747, 0, 32000],
            [],
        ),
    ],
)
84
@pytest.mark.parametrize("start_idx", [0, 4, 8])
85
# yapf: enable
86
87
88
def test_iter_token_matches(token_ids, match_ids, expected, start_idx):
    result = list(iter_token_matches(token_ids, match_ids,
                                     start_idx=start_idx))
89
90

    # Manually constructed results
91
92
    assert [item._asdict() for item in result
            ] == [item for item in expected if item["start_idx"] >= start_idx]
93
94
95
96
97
98
99

    # Invariants
    match_lens = [end - start for start, end in result]
    print("match_lens:", match_lens)  # Only displayed on error
    assert all(match_len == len(match_ids) for match_len in match_lens)


100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
# yapf: disable
@pytest.mark.parametrize(
    ("token_ids", "match_ids", "new_ids", "expected"),
    [
        ([], [], [-1], []),
        ([], [32000], [-1], []),
        (
            [32000, 32000, 32000],
            [32000],
            [-1],
            [-1, -1, -1],
        ),
        (
            [32000, 32000, 32000],
            [32000, 32000],
            [-1],
            [-1, 32000],
        ),
        (
            [32000, 32000, 32000],
            [32000, 32000, 32000],
            [-1],
            [-1],
        ),
        (
            [9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918],
            [28747, 32000],
            [-1],
            [9833, -1, 32000, 32000, 9833, -1, 32000, 918],
        ),
        (
            [9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918],
            [28747, 32000, 32000, 32000],
            [-1],
            [9833, -1, 9833, 28747, 32000, 32000, 918],
        ),
        (
            [9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918],
            [28747, 0, 32000],
            [-1],
            [9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918],
        ),
    ],
)
# yapf: enable
def test_replace_token_matches(token_ids, match_ids, new_ids, expected):
    result = replace_token_matches(token_ids, match_ids, new_ids)

    # Manually constructed results
    assert result == expected


152
153
154
155
156
157
158
159
160
# yapf: disable
@pytest.mark.parametrize(
    ("prompt", "target_by_key", "expected_by_key"),
    [
        (
            [],
            {
                "pattern_1": [],
                "pattern_2": [32000],
161
162
163
                "pattern_3": PromptIndexTargets.start(),
                "pattern_4": PromptIndexTargets.prefix([32000]),
                "pattern_5": PromptIndexTargets.end(),
164
165
            },
            {
166
                "pattern_1": [],
167
                "pattern_2": [],
168
169
170
171
172
173
174
                "pattern_3": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_4": [],
                "pattern_5": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
175
            },
176
177
178
179
180
181
182
        ),
        (
            [32000, 32000, 32000, 32000],
            {
                "pattern_1": [32000],
                "pattern_2": [32000, 32000],
                "pattern_3": [32000, 32000, 32000],
183
184
185
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix([32000]),
                "pattern_6": PromptIndexTargets.end(),
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
            },
            {
                "pattern_1": [
                    { "start_idx": 0, "end_idx": 1 },
                    { "start_idx": 1, "end_idx": 2 },
                    { "start_idx": 2, "end_idx": 3 },
                    { "start_idx": 3, "end_idx": 4 },
                ],
                "pattern_2": [
                    { "start_idx": 0, "end_idx": 2 },
                    { "start_idx": 2, "end_idx": 4 },
                ],
                "pattern_3": [
                    { "start_idx": 0, "end_idx": 3 },
                ],
201
202
203
204
205
206
207
208
209
                "pattern_4": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_5": [
                    { "start_idx": 1, "end_idx": 1 },
                ],
                "pattern_6": [
                    { "start_idx": 4, "end_idx": 4 },
                ],
210
211
212
213
214
215
216
217
            },
        ),
        (
            [9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918],
            {
                "pattern_1": [28747, 32000],
                "pattern_2": [28747, 32000, 32000, 32000],
                "pattern_3": [28747, 0, 32000],
218
219
220
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix([28747, 32000]),
                "pattern_6": PromptIndexTargets.end(),
221
222
223
224
225
226
227
228
229
230
            },
            {
                "pattern_1": [
                    { "start_idx": 1, "end_idx": 3 },
                    { "start_idx": 6, "end_idx": 8 },
                ],
                "pattern_2": [
                    { "start_idx": 1, "end_idx": 5 },
                ],
                "pattern_3": [],
231
232
233
234
235
236
237
                "pattern_4": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_5": [],
                "pattern_6": [
                    { "start_idx": 10, "end_idx": 10 },
                ],
238
239
240
241
            },
        ),
    ],
)
242
@pytest.mark.parametrize("update_type", [PromptInsertion, PromptReplacement])
243
# yapf: enable
244
245
246
247
248
249
def test_find_token_matches(
    prompt,
    target_by_key,
    expected_by_key,
    update_type,
):
250
251
252
    # Should not be used since there is nothing to convert to token IDs
    mock_tokenizer = cast(AnyTokenizer, object())

253
    prompt_updates = {
254
        key: update_type(key, target, []).resolve(0)
255
        for key, target in target_by_key.items()
256
257
258
259
260
    }
    result = {
        key: list(update.iter_token_matches(prompt, mock_tokenizer))
        for key, update in prompt_updates.items()
    }
261
262
263
264
265
266
267
268

    # Only displayed on error
    print("result:", result)

    # Manually constructed results
    assert {
        key: [
            dict(start_idx=item.start_idx, end_idx=item.end_idx)
269
            for item in result.get(key, [])
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
        ]
        for key in expected_by_key
    } == expected_by_key


# yapf: disable
@pytest.mark.parametrize(
    ("prompt", "target_by_key", "expected_by_key"),
    [
        # Detokenized test cases of `test_find_token_matches`
        # using the vocab of llava-hf/llava-v1.6-mistral-7b-hf
        (
            "",
            {
                "pattern_1": "",
                "pattern_2": "<image>",
286
287
288
                "pattern_3": PromptIndexTargets.start(),
                "pattern_4": PromptIndexTargets.prefix("<image>"),
                "pattern_5": PromptIndexTargets.end(),
289
290
291
292
            },
            {
                "pattern_1": [{ "start_idx": 0, "end_idx": 0 }],
                "pattern_2": [],
293
294
295
296
297
298
299
                "pattern_3": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_4": [],
                "pattern_5": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
300
301
302
303
304
305
306
307
            }
        ),
        (
            "<image><image><image><image>",
            {
                "pattern_1": "<image>",
                "pattern_2": "<image><image>",
                "pattern_3": "<image><image><image>",
308
309
310
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix("<image>"),
                "pattern_6": PromptIndexTargets.end(),
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
            },
            {
                "pattern_1": [
                    { "start_idx": 0, "end_idx": 7 },
                    { "start_idx": 7, "end_idx": 14 },
                    { "start_idx": 14, "end_idx": 21 },
                    { "start_idx": 21, "end_idx": 28 },
                ],
                "pattern_2": [
                    { "start_idx": 0, "end_idx": 14 },
                    { "start_idx": 14, "end_idx": 28 },
                ],
                "pattern_3": [
                    { "start_idx": 0, "end_idx": 21 },
                ],
326
327
328
329
330
331
332
333
334
                "pattern_4": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_5": [
                    { "start_idx": 7, "end_idx": 7 },
                ],
                "pattern_6": [
                    { "start_idx": 28, "end_idx": 28 },
                ],
335
336
337
338
339
340
341
342
            },
        ),
        (
            "Image:<image><image><image>Image:<image><image>!",
            {
                "pattern_1": "Image:<image>",
                "pattern_2": "Image:<image><image><image>",
                "pattern_3": "Image:<unk><image>",
343
344
345
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix("Image:<image>"),
                "pattern_6": PromptIndexTargets.end(),
346
347
348
349
350
351
352
353
354
355
            },
            {
                "pattern_1": [
                    { "start_idx": 0, "end_idx": 13 },
                    { "start_idx": 27, "end_idx": 40 },
                ],
                "pattern_2": [
                    { "start_idx": 0, "end_idx": 27 },
                ],
                "pattern_3": [],
356
357
358
359
360
361
362
363
364
                "pattern_4": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_5": [
                    { "start_idx": 13, "end_idx": 13 },
                ],
                "pattern_6": [
                    { "start_idx": 48, "end_idx": 48 },
                ],
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
            },
        ),
        # Test regex escape
        (
            "<|image|><image><|image|><image>",
            {
                "pattern_1": "<|image|>",
                "pattern_2": "<|image|><image>",
                "pattern_3": "<|image|><image><|image|>",
            },
            {
                "pattern_1": [
                    { "start_idx": 0, "end_idx": 9 },
                    { "start_idx": 16, "end_idx": 25 },
                ],
                "pattern_2": [
                    { "start_idx": 0, "end_idx": 16 },
                    { "start_idx": 16, "end_idx": 32 },
                ],
                "pattern_3": [
                    { "start_idx": 0, "end_idx": 25 },
                ],
            },
        ),
    ],
)
391
@pytest.mark.parametrize("update_type", [PromptInsertion, PromptReplacement])
392
# yapf: enable
393
394
395
396
397
398
def test_find_text_matches(
    prompt,
    target_by_key,
    expected_by_key,
    update_type,
):
399
400
401
    # Should not be used since there is nothing to convert to text
    mock_tokenizer = cast(AnyTokenizer, object())

402
    prompt_updates = {
403
        key: update_type(key, target, []).resolve(0)
404
        for key, target in target_by_key.items()
405
406
407
408
409
    }
    result = {
        key: list(update.iter_text_matches(prompt, mock_tokenizer))
        for key, update in prompt_updates.items()
    }
410
411
412
413
414
415
416
417

    # Only displayed on error
    print("result:", result)

    # Manually constructed results
    assert {
        key: [
            dict(start_idx=item.start_idx, end_idx=item.end_idx)
418
            for item in result.get(key, [])
419
420
421
422
423
424
425
        ]
        for key in expected_by_key
    } == expected_by_key


# yapf: disable
@pytest.mark.parametrize(
426
    ("prompt", "target_by_key", "repl_by_key", "expected_by_update_type_mm_count"),  # noqa: E501
427
428
429
430
431
432
433
434
435
436
437
    [
        (
            "Image:<image>Image:<image><image>!",
            {
                # We use `<image>` before `Image:` to test matches that
                # occur out of order
                "pattern_1": "<image>",
                "pattern_2": "Image:",
                "pattern_3": "!",
            },
            {
438
439
440
441
442
443
                # Test whether target is confused with replacement
                "pattern_1": "<image><image>",
                # Test empty replacement
                "pattern_2": "",
                # Test dynamic replacement (beyond the form of `unit * count`)
                "pattern_3": "?!?",
444
            },
445
446
447
448
449
450
451
452
453
454
455
456
            {
                PromptInsertion: {
                    0: "Image:<image>Image:<image><image>!",
                    1: "Image:<image><image><image>Image:<image><image>!?!?",
                    2: "Image:<image><image><image><image><image>Image:<image><image>!?!??!?",  # noqa: E501
                },
                PromptReplacement: {
                    0: "Image:<image>Image:<image><image>!",
                    1: "<image><image>Image:<image><image>?!?",
                    2: "<image><image><image><image><image>?!?",
                },
            },
457
        ),
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
        # Test index targets
        (
            "",
            {
                "pattern_1": PromptIndexTargets.start(),
                "pattern_2": PromptIndexTargets.prefix("<image>"),
                "pattern_3": PromptIndexTargets.end(),
            },
            {
                "pattern_1": "1",
                "pattern_2": "2",
                "pattern_3": "3",
            },
            {
                PromptInsertion: {
                    0: "",
                    1: "13",
                    2: "1133",
                },
                PromptReplacement: {
                    0: "",
                    1: "13",
                    2: "1133",
                },
            },
        ),
        (
            "<image>",
            {
                "pattern_1": PromptIndexTargets.start(),
                "pattern_2": PromptIndexTargets.prefix("<image>"),
                "pattern_3": PromptIndexTargets.end(),
            },
            {
                "pattern_1": "1",
                "pattern_2": "2",
                "pattern_3": "3",
            },
            {
                PromptInsertion: {
                    0: "<image>",
                    1: "1<image>23",
                    2: "11<image>2233",
                },
                PromptReplacement: {
                    0: "<image>",
                    1: "1<image>23",
                    2: "11<image>2233",
                },
            },
        ),
        # Test different replacement per item
        (
            "<image><image><image>",
            {
                "pattern_1": "<image>",
            },
            {
                "pattern_1": lambda idx: str(idx + 1),
            },
            {
                PromptInsertion: {
                    0: "<image><image><image>",
                    1: "<image>1<image><image>",
                    2: "<image>12<image><image>",
                },
                PromptReplacement: {
                    0: "<image><image><image>",
                    1: "1<image><image>",
                    2: "12<image>",
                },
            },
        ),
        (
            "<image><image><image>",
            {
                "pattern_1": PromptIndexTargets.prefix("<image>"),
            },
            {
                "pattern_1": lambda idx: str(idx + 1),
            },
            {
                PromptInsertion: {
                    0: "<image><image><image>",
                    1: "<image>1<image><image>",
                    2: "<image>12<image><image>",
                },
                PromptReplacement: {
                    0: "<image><image><image>",
                    1: "<image>1<image><image>",
                    2: "<image>12<image><image>",
                },
            },
        ),
552
553
554
    ]
)
# yapf: enable
555
def test_find_update_text(
556
557
558
    prompt,
    target_by_key,
    repl_by_key,
559
    expected_by_update_type_mm_count,
560
561
562
563
):
    # Should not be used since there is nothing to convert to text
    mock_tokenizer = cast(AnyTokenizer, object())

564
565
566
567
568
    for (
            update_type,
            expected_by_mm_count,
    ) in expected_by_update_type_mm_count.items():
        for mm_count, expected in expected_by_mm_count.items():
569
            mm_prompt_updates = {
570
571
                key: [[update_type(key, target, repl_by_key[key]).resolve(i)]
                      for i in range(mm_count)]
572
573
574
575
                for key, target in target_by_key.items()
            }

            new_prompt, result = apply_text_matches(
576
                prompt,
577
578
                mm_prompt_updates,
                mock_tokenizer,
579
580
581
582
583
            )

            # Only displayed on error
            print("update_type:", update_type)
            print("mm_count:", mm_count)
584
585
            print("mm_prompt_updates:", mm_prompt_updates)
            print("new_prompt:", new_prompt)
586
587
588
            print("result:", result)

            # Manually constructed results
589
            assert new_prompt == expected
590
591
592
593


# yapf: disable
@pytest.mark.parametrize(
594
    ("prompt", "target_by_key", "repl_by_key", "expected_by_update_type_mm_count"),  # noqa: E501
595
    [
596
        # Tokenized test cases of `test_find_update_text`
597
598
599
600
601
602
603
604
605
606
607
        # using the vocab of llava-hf/llava-v1.6-mistral-7b-hf
        (
            [1, 9833, 28747, 32000, 9833, 28747, 32000, 32000, 918],
            {
                # We use `<image>` before `Image:` to test matches that
                # occur out of order
                "pattern_1": [32000],
                "pattern_2": [9833, 28747],
                "pattern_3": [918],
            },
            {
608
609
610
611
612
613
                # Test whether target is confused with replacement
                "pattern_1": [32000, 32000],
                # Test empty replacement
                "pattern_2": [],
                # Test dynamic replacement (beyond the form of `unit * count`)
                "pattern_3": [1550, 918, 1550],
614
            },
615
616
617
618
619
620
621
622
623
624
625
626
            {
                PromptInsertion: {
                    0: [1, 9833, 28747, 32000, 9833, 28747, 32000, 32000, 918],
                    1: [1, 9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918, 1550, 918, 1550],  # noqa: E501
                    2: [1, 9833, 28747, 32000, 32000, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918, 1550, 918, 1550, 1550, 918, 1550],  # noqa: E501
                },
                PromptReplacement: {
                    0: [1, 9833, 28747, 32000, 9833, 28747, 32000, 32000, 918],
                    1: [1, 32000, 32000, 9833, 28747, 32000, 32000, 1550, 918, 1550],  # noqa: E501
                    2: [1, 32000, 32000, 32000, 32000, 32000, 1550, 918, 1550],
                },
            },
627
        ),
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
        # Test index targets
        (
            [],
            {
                "pattern_1": PromptIndexTargets.start(),
                "pattern_2": PromptIndexTargets.prefix([32000]),
                "pattern_3": PromptIndexTargets.end(),
            },
            {
                "pattern_1": [-1],
                "pattern_2": [-2],
                "pattern_3": [-3],
            },
            {
                PromptInsertion: {
                    0: [],
                    1: [-1, -3],
                    2: [-1, -1, -3, -3],
                },
                PromptReplacement: {
                    0: [],
                    1: [-1, -3],
                    2: [-1, -1, -3, -3],
                },
            },
        ),
        (
            [32000],
            {
                "pattern_1": PromptIndexTargets.start(),
                "pattern_2": PromptIndexTargets.prefix([32000]),
                "pattern_3": PromptIndexTargets.end(),
            },
            {
                "pattern_1": [-1],
                "pattern_2": [-2],
                "pattern_3": [-3],
            },
            {
                PromptInsertion: {
                    0: [32000],
                    1: [-1, 32000, -2, -3],
                    2: [-1, -1, 32000, -2, -2, -3, -3],
                },
                PromptReplacement: {
                    0: [32000],
                    1: [-1, 32000, -2, -3],
                    2: [-1, -1, 32000, -2, -2, -3, -3],
                },
            },
        ),
        # Test different replacement per item
        (
            [32000, 32000, 32000],
            {
                "pattern_1": [32000],
            },
            {
                "pattern_1": lambda idx: [-(idx + 1)],
            },
            {
                PromptInsertion: {
                    0: [32000, 32000, 32000],
                    1: [32000, -1, 32000, 32000],
                    2: [32000, -1, -2, 32000, 32000],
                },
                PromptReplacement: {
                    0: [32000, 32000, 32000],
                    1: [-1, 32000, 32000],
                    2: [-1, -2, 32000],
                },
            },
        ),
        (
            [32000, 32000, 32000],
            {
                "pattern_1": PromptIndexTargets.prefix([32000]),
            },
            {
                "pattern_1": lambda idx: [-(idx + 1)],
            },
            {
                PromptInsertion: {
                    0: [32000, 32000, 32000],
                    1: [32000, -1, 32000, 32000],
                    2: [32000, -1, -2, 32000, 32000],
                },
                PromptReplacement: {
                    0: [32000, 32000, 32000],
                    1: [32000, -1, 32000, 32000],
                    2: [32000, -1, -2, 32000, 32000],
                },
            },
        ),
722
723
724
    ]
)
# yapf: enable
725
def test_find_update_tokens(
726
727
728
    prompt,
    target_by_key,
    repl_by_key,
729
    expected_by_update_type_mm_count,
730
731
732
733
):
    # Should not be used since there is nothing to convert to tokens
    mock_tokenizer = cast(AnyTokenizer, object())

734
735
736
737
738
    for (
            update_type,
            expected_by_mm_count,
    ) in expected_by_update_type_mm_count.items():
        for mm_count, expected in expected_by_mm_count.items():
739
            mm_prompt_updates = {
740
741
                key: [[update_type(key, target, repl_by_key[key]).resolve(i)]
                      for i in range(mm_count)]
742
743
744
745
                for key, target in target_by_key.items()
            }

            new_prompt, result = apply_token_matches(
746
                prompt,
747
748
                mm_prompt_updates,
                mock_tokenizer,
749
750
751
752
753
            )

            # Only displayed on error
            print("update_type:", update_type)
            print("mm_count:", mm_count)
754
755
            print("mm_prompt_updates:", mm_prompt_updates)
            print("new_prompt:", new_prompt)
756
757
758
            print("result:", result)

            # Manually constructed results
759
            assert new_prompt == expected
760
761
762
763
764
765
766


# yapf: disable
@pytest.mark.parametrize(
    "repl_by_key",
    [
        {
767
768
769
            "pattern_1": [32000, 32000],
            "pattern_2": [],
            "pattern_3": [1550, 918, 1550],
770
771
            # Test different modalities having the same tokens (32000)
            "pattern_4": [32000],
772
773
774
775
776
777
778
779
        },
    ],
)
@pytest.mark.parametrize(
    ("prompt", "expected"),
    [
        (
            [1, 9833, 28747, 32000, 9833, 28747, 32000, 32000, 918],
780
781
            {
                "pattern_1": [
782
                    PlaceholderFeaturesInfo(
783
784
785
                        modality="pattern_1",
                        item_idx=0,
                        start_idx=6,
786
                        tokens=[32000, 32000],
787
                        is_embed=None,
788
789
                    ),
                ],
790
                "pattern_4": [
791
                    PlaceholderFeaturesInfo(
792
793
794
                        modality="pattern_4",
                        item_idx=0,
                        start_idx=3,
795
                        tokens=[32000],
796
                        is_embed=None,
797
798
                    ),
                ],
799
800
            }

801
802
        ),
        (
803
            [1, 32000, 32000, 9833, 28747, 32000, 32000, 1550, 918, 1550],
804
805
            {
                "pattern_1": [
806
                    PlaceholderFeaturesInfo(
807
808
809
                        modality="pattern_1",
                        item_idx=0,
                        start_idx=1,
810
                        tokens=[32000, 32000],
811
                        is_embed=None,
812
                    ),
813
                    PlaceholderFeaturesInfo(
814
815
816
                        modality="pattern_1",
                        item_idx=1,
                        start_idx=5,
817
                        tokens=[32000, 32000],
818
                        is_embed=None,
819
820
821
                    ),
                ],
                "pattern_3": [
822
                    PlaceholderFeaturesInfo(
823
824
825
                        modality="pattern_3",
                        item_idx=0,
                        start_idx=7,
826
                        tokens=[1550, 918, 1550],
827
                        is_embed=None,
828
829
                    ),
                ],
830
                # No match for pattern_4 as it has lower priority than pattern_1
831
            }
832
833
        ),
        (
834
            [1, 32000, 32000, 32000, 32000, 32000, 1550, 918, 1550],
835
836
            {
                "pattern_1": [
837
                    PlaceholderFeaturesInfo(
838
839
840
                        modality="pattern_1",
                        item_idx=0,
                        start_idx=1,
841
                        tokens=[32000, 32000],
842
                        is_embed=None,
843
                    ),
844
                    PlaceholderFeaturesInfo(
845
846
847
                        modality="pattern_1",
                        item_idx=1,
                        start_idx=3,
848
                        tokens=[32000, 32000],
849
                        is_embed=None,
850
851
                    ),
                ],
852
                "pattern_4": [
853
                    PlaceholderFeaturesInfo(
854
855
856
                        modality="pattern_4",
                        item_idx=0,
                        start_idx=5,
857
                        tokens=[32000],
858
                        is_embed=None,
859
860
                    ),
                ],
861
                "pattern_3": [
862
                    PlaceholderFeaturesInfo(
863
864
865
                        modality="pattern_3",
                        item_idx=0,
                        start_idx=6,
866
                        tokens=[1550, 918, 1550],
867
                        is_embed=None,
868
869
870
                    ),
                ],
            }
871
872
873
        ),
    ]
)
874
@pytest.mark.parametrize("update_type", [PromptInsertion, PromptReplacement])
875
# yapf: enable
876
def test_find_mm_placeholders(
877
878
879
    repl_by_key,
    prompt,
    expected,
880
    update_type,
881
882
883
884
):
    # Should not be used since there is nothing to convert to tokens
    mock_tokenizer = cast(AnyTokenizer, object())

885
    mm_prompt_updates = {
886
        key: [[update_type(key, [], repl).resolve(i)] for i in range(3)]
887
        for key, repl in repl_by_key.items()
888
    }
889

890
    result = find_mm_placeholders(prompt, mm_prompt_updates, mock_tokenizer)
891
892
893

    # Only displayed on error
    print("result:", result)
894
895

    # Manually constructed results
896
    assert result == expected
897
898


899
@pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
900
901
902
903
904
905
906
907
908
909
910
911
912
@pytest.mark.parametrize(
    ("limit", "num_supported", "is_valid"),
    [(0, 0, True), (0, 1, True), (1, 0, False), (1, 1, True), (1, 2, True),
     (2, 1, False), (2, 2, True)],
)
def test_limit_mm_per_prompt_dummy(model_id, limit, num_supported, is_valid):
    limit_mm_per_prompt = {"image": limit}

    model_config = ModelConfig(
        model=model_id,
        limit_mm_per_prompt=limit_mm_per_prompt,
    )

913
    processor = MULTIMODAL_REGISTRY.create_processor(model_config)
914
    processor._supported_mm_limits = {"image": num_supported}
915

916
    profiler = MultiModalProfiler(processor)
917
918
919
920

    if is_valid:
        exc_ctx = nullcontext()
    else:
921
        exc_ctx = pytest.raises(ValueError, match="At most")
922
923

    with exc_ctx:
924
925
926
927
        profiler.get_decoder_dummy_data(
            model_config.max_model_len,
            mm_counts=limit_mm_per_prompt,
        )
928
929


930
@pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
931
932
933
934
935
936
937
938
939
940
941
942
943
@pytest.mark.parametrize(
    ("num_images", "limit", "is_valid"),
    [(0, 0, True), (0, 1, True), (1, 0, False), (1, 1, True), (1, 2, True),
     (2, 1, False), (2, 2, True)],
)
def test_limit_mm_per_prompt_apply(model_id, num_images, limit, is_valid):
    limit_mm_per_prompt = {"image": limit}

    model_config = ModelConfig(
        model=model_id,
        limit_mm_per_prompt=limit_mm_per_prompt,
    )

944
    processor = MULTIMODAL_REGISTRY.create_processor(model_config)
945
946

    rng = np.random.RandomState(0)
947
    image = random_image(rng, min_wh=128, max_wh=256)
948
949
950
951
952
953
954
955
956
957
    if num_images == 0:
        mm_data = {}
    elif num_images == 1:
        mm_data = {"image": image}
    else:
        mm_data = {"image": [image] * num_images}

    if is_valid:
        exc_ctx = nullcontext()
    else:
958
        exc_ctx = pytest.raises(ValueError, match="At most")
959
960
961
962
963
964
965

    with exc_ctx:
        processor.apply(
            "<image>" * num_images,
            mm_data=mm_data,
            hf_processor_mm_kwargs={},
        )
966
967


968
class DummyProcessor:
969

970
    def __init__(self, a: int = 0, b: int = 0) -> None:
971
972
        super().__init__()

973
974
        self.a = a
        self.b = b
975
976
977

    def __call__(
        self,
978
979
980
981
982
        a: int = 0,
        c: int = 0,
        return_tensors: Optional[str] = None,
    ) -> dict[str, int]:
        return dict(a=a, c=c)
983
984
985


# yapf: disable
986
@pytest.mark.parametrize("model_id", ["Qwen/Qwen2-VL-2B-Instruct"])  # Dummy
987
@pytest.mark.parametrize(
988
    ("config_kwargs", "inference_kwargs", "expected_kwargs"),
989
    [
990
991
992
993
994
995
996
        ({"a": 1}, {}, {"a": 1, "b": 0}),
        ({}, {"a": 1}, {"a": 1, "b": 0}),
        # inference_kwargs should take precedence
        ({"a": 1}, {"a": 2}, {"a": 2, "b": 0}),
        # Should ignore extra kwargs
        ({"a": 1, "c": 1}, {}, {"a": 1, "b": 0}),
        ({"b": 1, "c": 1}, {}, {"a": 0, "b": 1}),
997
998
999
    ],
)
# yapf: enable
1000
1001
1002
1003
1004
1005
1006
1007
def test_hf_processor_init_kwargs(
    model_id,
    config_kwargs,
    inference_kwargs,
    expected_kwargs,
):
    # Should not be used since there is nothing to convert to tokens
    mock_tokenizer = cast(AnyTokenizer, object())
1008

1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
    ctx = InputProcessingContext(
        model_config=ModelConfig(model_id, mm_processor_kwargs=config_kwargs),
        tokenizer=mock_tokenizer,
    )

    processor = ctx.get_hf_processor(
        DummyProcessor,  # type: ignore[arg-type]
        **inference_kwargs,
    )

    for k, v in expected_kwargs.items():
        assert getattr(processor, k) == v
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
# yapf: disable
@pytest.mark.parametrize("model_id", ["Qwen/Qwen2-VL-2B-Instruct"])  # Dummy
@pytest.mark.parametrize(
    ("config_kwargs", "inference_kwargs", "expected_kwargs"),
    [
        ({"a": 1}, {}, {"a": 1, "c": 0}),
        ({}, {"a": 1}, {"a": 1, "c": 0}),
        # inference_kwargs should take precedence
        ({"a": 1}, {"a": 2}, {"a": 2, "c": 0}),
        # Should ignore extra kwargs
        ({"a": 1, "c": 1}, {}, {"a": 1, "c": 1}),
        ({"b": 1, "c": 1}, {}, {"a": 0, "c": 1}),
    ],
)
# yapf: enable
def test_hf_processor_call_kwargs(
    model_id,
    config_kwargs,
    inference_kwargs,
    expected_kwargs,
):
    # Should not be used since there is nothing to convert to tokens
    mock_tokenizer = cast(AnyTokenizer, object())
1046

1047
1048
1049
    ctx = InputProcessingContext(
        model_config=ModelConfig(model_id, mm_processor_kwargs=config_kwargs),
        tokenizer=mock_tokenizer,
1050
1051
    )

1052
1053
1054
1055
    processor = ctx.get_hf_processor(DummyProcessor)  # type: ignore[arg-type]

    result = ctx.call_hf_processor(processor, {}, inference_kwargs)
    assert result == expected_kwargs