test_processing.py 33.9 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import time
5
from contextlib import nullcontext
6
from typing import cast
7

8
import numpy as np
9
import pytest
10
11
12

from vllm.config import ModelConfig
from vllm.multimodal import MULTIMODAL_REGISTRY
13
14
15
16
17
18
from vllm.multimodal.processing import (
    InputProcessingContext,
    PlaceholderFeaturesInfo,
    PromptIndexTargets,
    PromptInsertion,
    PromptReplacement,
19
    _apply_matches,
20
21
22
23
24
25
    apply_text_matches,
    apply_token_matches,
    find_mm_placeholders,
    iter_token_matches,
    replace_token_matches,
)
26
from vllm.multimodal.profiling import MultiModalProfiler
27
from vllm.tokenizers import TokenizerLike
28

29
30
from .utils import random_image

31
32
pytestmark = pytest.mark.cpu_test

33
34
35
36

@pytest.mark.parametrize(
    ("token_ids", "match_ids", "expected"),
    [
37
        ([], [], []),
38
39
40
41
42
        ([], [32000], []),
        (
            [32000, 32000, 32000],
            [32000],
            [
43
44
45
                {"start_idx": 0, "end_idx": 1},
                {"start_idx": 1, "end_idx": 2},
                {"start_idx": 2, "end_idx": 3},
46
47
48
49
50
            ],
        ),
        (
            [32000, 32000, 32000],
            [32000, 32000],
51
            [{"start_idx": 0, "end_idx": 2}],
52
53
54
55
        ),
        (
            [32000, 32000, 32000],
            [32000, 32000, 32000],
56
            [{"start_idx": 0, "end_idx": 3}],
57
58
59
60
61
        ),
        (
            [9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918],
            [28747, 32000],
            [
62
63
                {"start_idx": 1, "end_idx": 3},
                {"start_idx": 6, "end_idx": 8},
64
65
66
67
68
69
            ],
        ),
        (
            [9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918],
            [28747, 32000, 32000, 32000],
            [
70
                {"start_idx": 1, "end_idx": 5},
71
72
73
74
75
76
77
78
79
            ],
        ),
        (
            [9833, 28747, 32000, 32000, 32000, 9833, 28747, 32000, 32000, 918],
            [28747, 0, 32000],
            [],
        ),
    ],
)
80
81
@pytest.mark.parametrize("start_idx", [0, 4, 8])
def test_iter_token_matches(token_ids, match_ids, expected, start_idx):
82
    result = list(iter_token_matches(token_ids, match_ids, start_idx=start_idx))
83
84

    # Manually constructed results
85
86
87
    assert [item._asdict() for item in result] == [
        item for item in expected if item["start_idx"] >= start_idx
    ]
88
89
90
91
92
93
94

    # 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)


95
96
97
98
99
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
@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],
        ),
    ],
)
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


145
146
147
148
149
150
151
152
@pytest.mark.parametrize(
    ("prompt", "target_by_key", "expected_by_key"),
    [
        (
            [],
            {
                "pattern_1": [],
                "pattern_2": [32000],
153
154
155
                "pattern_3": PromptIndexTargets.start(),
                "pattern_4": PromptIndexTargets.prefix([32000]),
                "pattern_5": PromptIndexTargets.end(),
156
157
            },
            {
158
                "pattern_1": [],
159
                "pattern_2": [],
160
                "pattern_3": [
161
                    {"start_idx": 0, "end_idx": 0},
162
163
164
                ],
                "pattern_4": [],
                "pattern_5": [
165
                    {"start_idx": 0, "end_idx": 0},
166
                ],
167
            },
168
169
170
171
172
173
174
        ),
        (
            [32000, 32000, 32000, 32000],
            {
                "pattern_1": [32000],
                "pattern_2": [32000, 32000],
                "pattern_3": [32000, 32000, 32000],
175
176
177
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix([32000]),
                "pattern_6": PromptIndexTargets.end(),
178
179
180
            },
            {
                "pattern_1": [
181
182
183
184
                    {"start_idx": 0, "end_idx": 1},
                    {"start_idx": 1, "end_idx": 2},
                    {"start_idx": 2, "end_idx": 3},
                    {"start_idx": 3, "end_idx": 4},
185
186
                ],
                "pattern_2": [
187
188
                    {"start_idx": 0, "end_idx": 2},
                    {"start_idx": 2, "end_idx": 4},
189
190
                ],
                "pattern_3": [
191
                    {"start_idx": 0, "end_idx": 3},
192
                ],
193
                "pattern_4": [
194
                    {"start_idx": 0, "end_idx": 0},
195
196
                ],
                "pattern_5": [
197
                    {"start_idx": 1, "end_idx": 1},
198
199
                ],
                "pattern_6": [
200
                    {"start_idx": 4, "end_idx": 4},
201
                ],
202
203
204
205
206
207
208
209
            },
        ),
        (
            [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],
210
211
212
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix([28747, 32000]),
                "pattern_6": PromptIndexTargets.end(),
213
214
215
            },
            {
                "pattern_1": [
216
217
                    {"start_idx": 1, "end_idx": 3},
                    {"start_idx": 6, "end_idx": 8},
218
219
                ],
                "pattern_2": [
220
                    {"start_idx": 1, "end_idx": 5},
221
222
                ],
                "pattern_3": [],
223
                "pattern_4": [
224
                    {"start_idx": 0, "end_idx": 0},
225
226
227
                ],
                "pattern_5": [],
                "pattern_6": [
228
                    {"start_idx": 10, "end_idx": 10},
229
                ],
230
231
232
233
            },
        ),
    ],
)
234
235
236
237
238
239
240
@pytest.mark.parametrize("update_type", [PromptInsertion, PromptReplacement])
def test_find_token_matches(
    prompt,
    target_by_key,
    expected_by_key,
    update_type,
):
241
    # Should not be used since there is nothing to convert to token IDs
242
    mock_tokenizer = cast(TokenizerLike, object())
243

244
    prompt_updates = {
245
        key: update_type(key, target, []).resolve(0)
246
        for key, target in target_by_key.items()
247
248
249
250
251
    }
    result = {
        key: list(update.iter_token_matches(prompt, mock_tokenizer))
        for key, update in prompt_updates.items()
    }
252
253
254
255
256
257
258
259

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

    # Manually constructed results
    assert {
        key: [
            dict(start_idx=item.start_idx, end_idx=item.end_idx)
260
            for item in result.get(key, [])
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
        ]
        for key in expected_by_key
    } == expected_by_key


@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>",
276
277
278
                "pattern_3": PromptIndexTargets.start(),
                "pattern_4": PromptIndexTargets.prefix("<image>"),
                "pattern_5": PromptIndexTargets.end(),
279
280
            },
            {
281
                "pattern_1": [{"start_idx": 0, "end_idx": 0}],
282
                "pattern_2": [],
283
                "pattern_3": [
284
                    {"start_idx": 0, "end_idx": 0},
285
286
287
                ],
                "pattern_4": [],
                "pattern_5": [
288
                    {"start_idx": 0, "end_idx": 0},
289
                ],
290
            },
291
292
293
294
295
296
297
        ),
        (
            "<image><image><image><image>",
            {
                "pattern_1": "<image>",
                "pattern_2": "<image><image>",
                "pattern_3": "<image><image><image>",
298
299
300
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix("<image>"),
                "pattern_6": PromptIndexTargets.end(),
301
302
303
            },
            {
                "pattern_1": [
304
305
306
307
                    {"start_idx": 0, "end_idx": 7},
                    {"start_idx": 7, "end_idx": 14},
                    {"start_idx": 14, "end_idx": 21},
                    {"start_idx": 21, "end_idx": 28},
308
309
                ],
                "pattern_2": [
310
311
                    {"start_idx": 0, "end_idx": 14},
                    {"start_idx": 14, "end_idx": 28},
312
313
                ],
                "pattern_3": [
314
                    {"start_idx": 0, "end_idx": 21},
315
                ],
316
                "pattern_4": [
317
                    {"start_idx": 0, "end_idx": 0},
318
319
                ],
                "pattern_5": [
320
                    {"start_idx": 7, "end_idx": 7},
321
322
                ],
                "pattern_6": [
323
                    {"start_idx": 28, "end_idx": 28},
324
                ],
325
326
327
328
329
330
331
332
            },
        ),
        (
            "Image:<image><image><image>Image:<image><image>!",
            {
                "pattern_1": "Image:<image>",
                "pattern_2": "Image:<image><image><image>",
                "pattern_3": "Image:<unk><image>",
333
334
335
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix("Image:<image>"),
                "pattern_6": PromptIndexTargets.end(),
336
337
338
            },
            {
                "pattern_1": [
339
340
                    {"start_idx": 0, "end_idx": 13},
                    {"start_idx": 27, "end_idx": 40},
341
342
                ],
                "pattern_2": [
343
                    {"start_idx": 0, "end_idx": 27},
344
345
                ],
                "pattern_3": [],
346
                "pattern_4": [
347
                    {"start_idx": 0, "end_idx": 0},
348
349
                ],
                "pattern_5": [
350
                    {"start_idx": 13, "end_idx": 13},
351
352
                ],
                "pattern_6": [
353
                    {"start_idx": 48, "end_idx": 48},
354
                ],
355
356
357
358
359
360
361
362
363
364
365
366
            },
        ),
        # Test regex escape
        (
            "<|image|><image><|image|><image>",
            {
                "pattern_1": "<|image|>",
                "pattern_2": "<|image|><image>",
                "pattern_3": "<|image|><image><|image|>",
            },
            {
                "pattern_1": [
367
368
                    {"start_idx": 0, "end_idx": 9},
                    {"start_idx": 16, "end_idx": 25},
369
370
                ],
                "pattern_2": [
371
372
                    {"start_idx": 0, "end_idx": 16},
                    {"start_idx": 16, "end_idx": 32},
373
374
                ],
                "pattern_3": [
375
                    {"start_idx": 0, "end_idx": 25},
376
377
378
379
380
                ],
            },
        ),
    ],
)
381
382
383
384
385
386
387
@pytest.mark.parametrize("update_type", [PromptInsertion, PromptReplacement])
def test_find_text_matches(
    prompt,
    target_by_key,
    expected_by_key,
    update_type,
):
388
    # Should not be used since there is nothing to convert to text
389
    mock_tokenizer = cast(TokenizerLike, object())
390

391
    prompt_updates = {
392
        key: update_type(key, target, []).resolve(0)
393
        for key, target in target_by_key.items()
394
395
396
397
398
    }
    result = {
        key: list(update.iter_text_matches(prompt, mock_tokenizer))
        for key, update in prompt_updates.items()
    }
399
400
401
402
403
404
405
406

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

    # Manually constructed results
    assert {
        key: [
            dict(start_idx=item.start_idx, end_idx=item.end_idx)
407
            for item in result.get(key, [])
408
409
410
411
412
413
        ]
        for key in expected_by_key
    } == expected_by_key


@pytest.mark.parametrize(
414
    ("prompt", "target_by_key", "repl_by_key", "expected_by_update_type_mm_count"),  # noqa: E501
415
416
417
418
419
420
421
422
423
424
425
    [
        (
            "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": "!",
            },
            {
426
427
428
429
430
431
                # 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": "?!?",
432
            },
433
434
435
436
437
438
439
440
441
442
443
444
            {
                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>?!?",
                },
            },
445
        ),
446
447
448
449
450
451
452
453
454
455
456
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
        # 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>",
                },
            },
        ),
540
    ],
541
)
542
def test_find_update_text(
543
544
545
    prompt,
    target_by_key,
    repl_by_key,
546
    expected_by_update_type_mm_count,
547
548
):
    # Should not be used since there is nothing to convert to text
549
    mock_tokenizer = cast(TokenizerLike, object())
550

551
    for (
552
553
        update_type,
        expected_by_mm_count,
554
555
    ) in expected_by_update_type_mm_count.items():
        for mm_count, expected in expected_by_mm_count.items():
556
            mm_prompt_updates = {
557
558
559
560
                key: [
                    [update_type(key, target, repl_by_key[key]).resolve(i)]
                    for i in range(mm_count)
                ]
561
562
563
564
                for key, target in target_by_key.items()
            }

            new_prompt, result = apply_text_matches(
565
                prompt,
566
567
                mm_prompt_updates,
                mock_tokenizer,
568
569
570
571
572
            )

            # Only displayed on error
            print("update_type:", update_type)
            print("mm_count:", mm_count)
573
574
            print("mm_prompt_updates:", mm_prompt_updates)
            print("new_prompt:", new_prompt)
575
576
577
            print("result:", result)

            # Manually constructed results
578
            assert new_prompt == expected
579
580
581


@pytest.mark.parametrize(
582
    ("prompt", "target_by_key", "repl_by_key", "expected_by_update_type_mm_count"),  # noqa: E501
583
    [
584
        # Tokenized test cases of `test_find_update_text`
585
586
587
588
589
590
591
592
593
594
595
        # 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],
            },
            {
596
597
598
599
600
601
                # 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],
602
            },
603
604
605
            {
                PromptInsertion: {
                    0: [1, 9833, 28747, 32000, 9833, 28747, 32000, 32000, 918],
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
                    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
643
644
645
646
647
648
649
                },
                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],
                },
            },
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
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
        # 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],
                },
            },
        ),
745
    ],
746
)
747
def test_find_update_tokens(
748
749
750
    prompt,
    target_by_key,
    repl_by_key,
751
    expected_by_update_type_mm_count,
752
753
):
    # Should not be used since there is nothing to convert to tokens
754
    mock_tokenizer = cast(TokenizerLike, object())
755

756
    for (
757
758
        update_type,
        expected_by_mm_count,
759
760
    ) in expected_by_update_type_mm_count.items():
        for mm_count, expected in expected_by_mm_count.items():
761
            mm_prompt_updates = {
762
763
764
765
                key: [
                    [update_type(key, target, repl_by_key[key]).resolve(i)]
                    for i in range(mm_count)
                ]
766
767
768
769
                for key, target in target_by_key.items()
            }

            new_prompt, result = apply_token_matches(
770
                prompt,
771
772
                mm_prompt_updates,
                mock_tokenizer,
773
774
775
776
777
            )

            # Only displayed on error
            print("update_type:", update_type)
            print("mm_count:", mm_count)
778
779
            print("mm_prompt_updates:", mm_prompt_updates)
            print("new_prompt:", new_prompt)
780
781
782
            print("result:", result)

            # Manually constructed results
783
            assert new_prompt == expected
784
785
786
787
788
789


@pytest.mark.parametrize(
    "repl_by_key",
    [
        {
790
791
792
            "pattern_1": [32000, 32000],
            "pattern_2": [],
            "pattern_3": [1550, 918, 1550],
793
794
            # Test different modalities having the same tokens (32000)
            "pattern_4": [32000],
795
796
797
798
799
800
801
802
        },
    ],
)
@pytest.mark.parametrize(
    ("prompt", "expected"),
    [
        (
            [1, 9833, 28747, 32000, 9833, 28747, 32000, 32000, 918],
803
804
            {
                "pattern_1": [
805
                    PlaceholderFeaturesInfo(
806
807
808
                        modality="pattern_1",
                        item_idx=0,
                        start_idx=6,
809
                        tokens=[32000, 32000],
810
                        is_embed=None,
811
812
                    ),
                ],
813
                "pattern_4": [
814
                    PlaceholderFeaturesInfo(
815
816
817
                        modality="pattern_4",
                        item_idx=0,
                        start_idx=3,
818
                        tokens=[32000],
819
                        is_embed=None,
820
821
                    ),
                ],
822
            },
823
824
        ),
        (
825
            [1, 32000, 32000, 9833, 28747, 32000, 32000, 1550, 918, 1550],
826
827
            {
                "pattern_1": [
828
                    PlaceholderFeaturesInfo(
829
830
831
                        modality="pattern_1",
                        item_idx=0,
                        start_idx=1,
832
                        tokens=[32000, 32000],
833
                        is_embed=None,
834
                    ),
835
                    PlaceholderFeaturesInfo(
836
837
838
                        modality="pattern_1",
                        item_idx=1,
                        start_idx=5,
839
                        tokens=[32000, 32000],
840
                        is_embed=None,
841
842
843
                    ),
                ],
                "pattern_3": [
844
                    PlaceholderFeaturesInfo(
845
846
847
                        modality="pattern_3",
                        item_idx=0,
                        start_idx=7,
848
                        tokens=[1550, 918, 1550],
849
                        is_embed=None,
850
851
                    ),
                ],
852
                # No match for pattern_4 as it has lower priority than pattern_1
853
            },
854
855
        ),
        (
856
            [1, 32000, 32000, 32000, 32000, 32000, 1550, 918, 1550],
857
858
            {
                "pattern_1": [
859
                    PlaceholderFeaturesInfo(
860
861
862
                        modality="pattern_1",
                        item_idx=0,
                        start_idx=1,
863
                        tokens=[32000, 32000],
864
                        is_embed=None,
865
                    ),
866
                    PlaceholderFeaturesInfo(
867
868
869
                        modality="pattern_1",
                        item_idx=1,
                        start_idx=3,
870
                        tokens=[32000, 32000],
871
                        is_embed=None,
872
873
                    ),
                ],
874
                "pattern_4": [
875
                    PlaceholderFeaturesInfo(
876
877
878
                        modality="pattern_4",
                        item_idx=0,
                        start_idx=5,
879
                        tokens=[32000],
880
                        is_embed=None,
881
882
                    ),
                ],
883
                "pattern_3": [
884
                    PlaceholderFeaturesInfo(
885
886
887
                        modality="pattern_3",
                        item_idx=0,
                        start_idx=6,
888
                        tokens=[1550, 918, 1550],
889
                        is_embed=None,
890
891
                    ),
                ],
892
            },
893
        ),
894
    ],
895
)
896
@pytest.mark.parametrize("update_type", [PromptInsertion, PromptReplacement])
897
def test_find_mm_placeholders(
898
899
900
    repl_by_key,
    prompt,
    expected,
901
    update_type,
902
903
):
    # Should not be used since there is nothing to convert to tokens
904
    mock_tokenizer = cast(TokenizerLike, object())
905

906
    mm_prompt_updates = {
907
        key: [[update_type(key, [], repl).resolve(i)] for i in range(3)]
908
        for key, repl in repl_by_key.items()
909
    }
910

911
    result = find_mm_placeholders(prompt, mm_prompt_updates, mock_tokenizer)
912
913
914

    # Only displayed on error
    print("result:", result)
915
916

    # Manually constructed results
917
    assert result == expected
918
919


920
@pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
921
922
@pytest.mark.parametrize(
    ("limit", "num_supported", "is_valid"),
923
924
925
926
927
928
929
930
931
    [
        (0, 0, True),
        (0, 1, True),
        (1, 0, False),
        (1, 1, True),
        (1, 2, True),
        (2, 1, False),
        (2, 2, True),
    ],
932
933
934
935
936
937
938
939
940
)
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,
    )

941
    processor = MULTIMODAL_REGISTRY.create_processor(model_config)
942
    processor._supported_mm_limits = {"image": num_supported}
943

944
    profiler = MultiModalProfiler(processor)
945

946
    exc_ctx = nullcontext() if is_valid else pytest.raises(ValueError, match="At most")
947
948

    with exc_ctx:
949
950
951
952
        profiler.get_decoder_dummy_data(
            model_config.max_model_len,
            mm_counts=limit_mm_per_prompt,
        )
953
954


955
@pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
956
957
@pytest.mark.parametrize(
    ("num_images", "limit", "is_valid"),
958
959
960
961
962
963
964
965
966
    [
        (0, 0, True),
        (0, 1, True),
        (1, 0, False),
        (1, 1, True),
        (1, 2, True),
        (2, 1, False),
        (2, 2, True),
    ],
967
968
969
970
971
972
973
974
975
)
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,
    )

976
    processor = MULTIMODAL_REGISTRY.create_processor(model_config)
977
978

    rng = np.random.RandomState(0)
979
    image = random_image(rng, min_wh=128, max_wh=256)
980
981
982
983
984
985
986
    if num_images == 0:
        mm_data = {}
    elif num_images == 1:
        mm_data = {"image": image}
    else:
        mm_data = {"image": [image] * num_images}

987
    exc_ctx = nullcontext() if is_valid else pytest.raises(ValueError, match="At most")
988
989
990
991
992
993
994

    with exc_ctx:
        processor.apply(
            "<image>" * num_images,
            mm_data=mm_data,
            hf_processor_mm_kwargs={},
        )
995
996


997
998
class DummyProcessor:
    def __init__(self, a: int = 0, b: int = 0) -> None:
999
1000
        super().__init__()

1001
1002
        self.a = a
        self.b = b
1003
1004
1005

    def __call__(
        self,
1006
1007
        a: int = 0,
        c: int = 0,
1008
        return_tensors: str | None = None,
1009
1010
    ) -> dict[str, int]:
        return dict(a=a, c=c)
1011
1012


1013
@pytest.mark.parametrize("model_id", ["Qwen/Qwen2-VL-2B-Instruct"])  # Dummy
1014
@pytest.mark.parametrize(
1015
    ("config_kwargs", "inference_kwargs", "expected_kwargs"),
1016
    [
1017
1018
1019
1020
1021
1022
1023
        ({"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}),
1024
1025
    ],
)
1026
1027
1028
1029
1030
1031
1032
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
1033
    mock_tokenizer = cast(TokenizerLike, object())
1034

1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
    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
1047
1048


1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
@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}),
    ],
)
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
1069
    mock_tokenizer = cast(TokenizerLike, object())
1070

1071
1072
1073
    ctx = InputProcessingContext(
        model_config=ModelConfig(model_id, mm_processor_kwargs=config_kwargs),
        tokenizer=mock_tokenizer,
1074
1075
    )

1076
1077
1078
1079
    processor = ctx.get_hf_processor(DummyProcessor)  # type: ignore[arg-type]

    result = ctx.call_hf_processor(processor, {}, inference_kwargs)
    assert result == expected_kwargs
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091


def test_apply_matches_no_match_exits_quickly():
    """
    Test that _apply_matches exits quickly when no matches are found.

    Previously, _apply_matches had O(n²) behavior when no match was found
    because it would increment start_idx by 1 each iteration while
    re-scanning the entire prompt from prev_end_idx=0.

    With the fix, it should exit immediately when no match is found.
    """
1092
    mock_tokenizer = cast(TokenizerLike, object())
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112

    # Create a long prompt with no placeholder
    long_prompt = "x" * 10000

    # Create update looking for a placeholder that doesn't exist
    mm_prompt_updates = {
        "image": [[PromptReplacement("image", "<image>", "REPLACED").resolve(0)]]
    }

    start = time.perf_counter()
    result, _ = _apply_matches(
        long_prompt,
        mm_prompt_updates,
        mock_tokenizer,
    )
    elapsed = time.perf_counter() - start

    # Should complete in < 100ms (was taking seconds before the fix)
    assert elapsed < 0.1, f"_apply_matches took {elapsed:.2f}s, expected < 0.1s"
    assert "".join(result) == long_prompt