test_processing.py 34 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
import torch
10
11

from vllm.config import ModelConfig
12
from vllm.inputs import InputProcessingContext
13
from vllm.multimodal import MULTIMODAL_REGISTRY
14
15
16
from vllm.multimodal.inputs import (MultiModalFieldElem, MultiModalKwargs,
                                    MultiModalKwargsItem,
                                    MultiModalSharedField)
17
18
19
# yapf conflicts with isort for this block
# yapf: disable
from vllm.multimodal.processing import (PlaceholderFeaturesInfo,
20
21
22
                                        ProcessingCache, PromptIndexTargets,
                                        PromptInsertion, PromptReplacement,
                                        apply_text_matches,
23
                                        apply_token_matches,
24
                                        find_mm_placeholders,
25
                                        find_text_matches, find_token_matches,
26
27
                                        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
from vllm.utils import full_groupby

33
34
from .utils import random_image

35
36
37
38
39

# yapf: disable
@pytest.mark.parametrize(
    ("token_ids", "match_ids", "expected"),
    [
40
        ([], [], []),
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
84
85
86
87
88
89
90
91
92
93
94
95
        ([], [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],
            [],
        ),
    ],
)
# yapf: enable
def test_iter_token_matches(token_ids, match_ids, expected):
    result = list(iter_token_matches(token_ids, match_ids))

    # Manually constructed results
    assert [item._asdict() for item in result] == expected

    # 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)


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
145
146
147
# 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


148
149
150
151
152
153
154
155
156
# yapf: disable
@pytest.mark.parametrize(
    ("prompt", "target_by_key", "expected_by_key"),
    [
        (
            [],
            {
                "pattern_1": [],
                "pattern_2": [32000],
157
158
159
                "pattern_3": PromptIndexTargets.start(),
                "pattern_4": PromptIndexTargets.prefix([32000]),
                "pattern_5": PromptIndexTargets.end(),
160
161
            },
            {
162
                "pattern_1": [],
163
                "pattern_2": [],
164
165
166
167
168
169
170
                "pattern_3": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_4": [],
                "pattern_5": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
171
            },
172
173
174
175
176
177
178
        ),
        (
            [32000, 32000, 32000, 32000],
            {
                "pattern_1": [32000],
                "pattern_2": [32000, 32000],
                "pattern_3": [32000, 32000, 32000],
179
180
181
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix([32000]),
                "pattern_6": PromptIndexTargets.end(),
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
            },
            {
                "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 },
                ],
197
198
199
200
201
202
203
204
205
                "pattern_4": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_5": [
                    { "start_idx": 1, "end_idx": 1 },
                ],
                "pattern_6": [
                    { "start_idx": 4, "end_idx": 4 },
                ],
206
207
208
209
210
211
212
213
            },
        ),
        (
            [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],
214
215
216
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix([28747, 32000]),
                "pattern_6": PromptIndexTargets.end(),
217
218
219
220
221
222
223
224
225
226
            },
            {
                "pattern_1": [
                    { "start_idx": 1, "end_idx": 3 },
                    { "start_idx": 6, "end_idx": 8 },
                ],
                "pattern_2": [
                    { "start_idx": 1, "end_idx": 5 },
                ],
                "pattern_3": [],
227
228
229
230
231
232
233
                "pattern_4": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_5": [],
                "pattern_6": [
                    { "start_idx": 10, "end_idx": 10 },
                ],
234
235
236
237
            },
        ),
    ],
)
238
@pytest.mark.parametrize("update_type", [PromptInsertion, PromptReplacement])
239
# yapf: enable
240
241
242
243
244
245
def test_find_token_matches(
    prompt,
    target_by_key,
    expected_by_key,
    update_type,
):
246
247
248
    # Should not be used since there is nothing to convert to token IDs
    mock_tokenizer = cast(AnyTokenizer, object())

249
250
    prompt_updates = [
        update_type(key, target, []).bind(mock_tokenizer)
251
252
        for key, target in target_by_key.items()
    ]
253
    result = find_token_matches(prompt, prompt_updates)
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279

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

    # Manually constructed results
    result_groups = dict(full_groupby(result, key=lambda x: x.modality))
    assert {
        key: [
            dict(start_idx=item.start_idx, end_idx=item.end_idx)
            for item in result_groups.get(key, [])
        ]
        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>",
280
281
282
                "pattern_3": PromptIndexTargets.start(),
                "pattern_4": PromptIndexTargets.prefix("<image>"),
                "pattern_5": PromptIndexTargets.end(),
283
284
285
286
            },
            {
                "pattern_1": [{ "start_idx": 0, "end_idx": 0 }],
                "pattern_2": [],
287
288
289
290
291
292
293
                "pattern_3": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_4": [],
                "pattern_5": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
294
295
296
297
298
299
300
301
            }
        ),
        (
            "<image><image><image><image>",
            {
                "pattern_1": "<image>",
                "pattern_2": "<image><image>",
                "pattern_3": "<image><image><image>",
302
303
304
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix("<image>"),
                "pattern_6": PromptIndexTargets.end(),
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
            },
            {
                "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 },
                ],
320
321
322
323
324
325
326
327
328
                "pattern_4": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_5": [
                    { "start_idx": 7, "end_idx": 7 },
                ],
                "pattern_6": [
                    { "start_idx": 28, "end_idx": 28 },
                ],
329
330
331
332
333
334
335
336
            },
        ),
        (
            "Image:<image><image><image>Image:<image><image>!",
            {
                "pattern_1": "Image:<image>",
                "pattern_2": "Image:<image><image><image>",
                "pattern_3": "Image:<unk><image>",
337
338
339
                "pattern_4": PromptIndexTargets.start(),
                "pattern_5": PromptIndexTargets.prefix("Image:<image>"),
                "pattern_6": PromptIndexTargets.end(),
340
341
342
343
344
345
346
347
348
349
            },
            {
                "pattern_1": [
                    { "start_idx": 0, "end_idx": 13 },
                    { "start_idx": 27, "end_idx": 40 },
                ],
                "pattern_2": [
                    { "start_idx": 0, "end_idx": 27 },
                ],
                "pattern_3": [],
350
351
352
353
354
355
356
357
358
                "pattern_4": [
                    { "start_idx": 0, "end_idx": 0 },
                ],
                "pattern_5": [
                    { "start_idx": 13, "end_idx": 13 },
                ],
                "pattern_6": [
                    { "start_idx": 48, "end_idx": 48 },
                ],
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
            },
        ),
        # 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 },
                ],
            },
        ),
    ],
)
385
@pytest.mark.parametrize("update_type", [PromptInsertion, PromptReplacement])
386
# yapf: enable
387
388
389
390
391
392
def test_find_text_matches(
    prompt,
    target_by_key,
    expected_by_key,
    update_type,
):
393
394
395
    # Should not be used since there is nothing to convert to text
    mock_tokenizer = cast(AnyTokenizer, object())

396
397
    prompt_updates = [
        update_type(key, target, []).bind(mock_tokenizer)
398
399
        for key, target in target_by_key.items()
    ]
400
    result = find_text_matches(prompt, prompt_updates)
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417

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

    # Manually constructed results
    result_groups = dict(full_groupby(result, key=lambda x: x.modality))
    assert {
        key: [
            dict(start_idx=item.start_idx, end_idx=item.end_idx)
            for item in result_groups.get(key, [])
        ]
        for key in expected_by_key
    } == expected_by_key


# yapf: disable
@pytest.mark.parametrize(
418
    ("prompt", "target_by_key", "repl_by_key", "expected_by_update_type_mm_count"),  # noqa: E501
419
420
421
422
423
424
425
426
427
428
429
    [
        (
            "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": "!",
            },
            {
430
431
432
433
434
435
                # 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": "?!?",
436
            },
437
438
439
440
441
442
443
444
445
446
447
448
            {
                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>?!?",
                },
            },
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
540
541
542
543
        # 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>",
                },
            },
        ),
544
545
546
    ]
)
# yapf: enable
547
def test_find_update_text(
548
549
550
    prompt,
    target_by_key,
    repl_by_key,
551
    expected_by_update_type_mm_count,
552
553
554
555
):
    # Should not be used since there is nothing to convert to text
    mock_tokenizer = cast(AnyTokenizer, object())

556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
    for (
            update_type,
            expected_by_mm_count,
    ) in expected_by_update_type_mm_count.items():
        mm_prompt_updates = {
            key:
            [update_type(key, target, repl_by_key[key]).bind(mock_tokenizer)]
            for key, target in target_by_key.items()
        }
        mm_matches = {
            key: find_text_matches(prompt, updates)
            for key, updates in mm_prompt_updates.items()
        }

        for mm_count, expected in expected_by_mm_count.items():
            result = apply_text_matches(
                prompt,
                mm_matches,
                {key: mm_count
                 for key in repl_by_key},
            )

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

            # Manually constructed results
            assert result == expected
586
587
588
589


# yapf: disable
@pytest.mark.parametrize(
590
    ("prompt", "target_by_key", "repl_by_key", "expected_by_update_type_mm_count"),  # noqa: E501
591
592
593
594
595
596
597
598
599
600
601
602
603
    [
        # Tokenized test cases of `test_find_replace_text`
        # 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],
            },
            {
604
605
606
607
608
609
                # 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],
610
            },
611
612
613
614
615
616
617
618
619
620
621
622
            {
                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],
                },
            },
623
        ),
624
625
626
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
        # 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],
                },
            },
        ),
718
719
720
    ]
)
# yapf: enable
721
def test_find_update_tokens(
722
723
724
    prompt,
    target_by_key,
    repl_by_key,
725
    expected_by_update_type_mm_count,
726
727
728
729
):
    # Should not be used since there is nothing to convert to tokens
    mock_tokenizer = cast(AnyTokenizer, object())

730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
    for (
            update_type,
            expected_by_mm_count,
    ) in expected_by_update_type_mm_count.items():
        mm_prompt_updates = {
            key:
            [update_type(key, target, repl_by_key[key]).bind(mock_tokenizer)]
            for key, target in target_by_key.items()
        }
        mm_matches = {
            key: find_token_matches(prompt, updates)
            for key, updates in mm_prompt_updates.items()
        }

        for mm_count, expected in expected_by_mm_count.items():
            result = apply_token_matches(
                prompt,
                mm_matches,
                {key: mm_count
                 for key in repl_by_key},
            )

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

            # Manually constructed results
            assert result == 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
886
    mm_prompt_updates = {
        key: [update_type(key, [], repl).bind(mock_tokenizer)]
887
        for key, repl in repl_by_key.items()
888
    }
889

890
    result = find_mm_placeholders(
891
        mm_prompt_updates,
892
893
894
895
896
        prompt,
        # Effectively match all occurrences in the prompt
        {key: 3
         for key in repl_by_key},
    )
897
898
899

    # Only displayed on error
    print("result:", result)
900
901

    # Manually constructed results
902
    assert result == expected
903
904


905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
def _dummy_elem(modality: str, key: str, size: int):
    return MultiModalFieldElem(
        modality=modality,
        key=key,
        data=torch.empty((size, ), dtype=torch.int8),
        field=MultiModalSharedField(1),
    )


def _dummy_item(modality: str, size_by_key: dict[str, int]):
    return MultiModalKwargsItem.from_elems([
        _dummy_elem(modality, key, size) for key, size in size_by_key.items()
    ])


def _dummy_kw(size_by_key_modality: dict[str, dict[str, int]]):
    return MultiModalKwargs.from_items([
        _dummy_item(modality, size_by_key)
        for modality, size_by_key in size_by_key_modality.items()
    ])


# yapf: disable
@pytest.mark.parametrize(
    ("item", "expected_size"),
    [
        (_dummy_item("a", {"a1": 100}), 100),
        (_dummy_item("a", {"a1": 100, "a2": 110}), 210),
        (_dummy_kw({"a": {"a1": 100, "a2": 110}, "b": {"b1": 120, "b2": 130}}), 460),  # noqa: E501
    ],
)
# yapf: enable
def test_cache_item_size(item, expected_size):
    cache = ProcessingCache.get_lru_cache(2048, type(item))
    cache[""] = item

    assert cache.currsize == expected_size


944
@pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
945
946
947
948
949
950
951
952
953
954
955
956
957
@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,
    )

958
    processor = MULTIMODAL_REGISTRY.create_processor(model_config)
959
    processor._supported_mm_limits = {"image": num_supported}
960

961
    profiler = MultiModalProfiler(processor)
962
963
964
965

    if is_valid:
        exc_ctx = nullcontext()
    else:
966
        exc_ctx = pytest.raises(ValueError, match="At most")
967
968

    with exc_ctx:
969
970
971
972
        profiler.get_decoder_dummy_data(
            model_config.max_model_len,
            mm_counts=limit_mm_per_prompt,
        )
973
974


975
@pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
976
977
978
979
980
981
982
983
984
985
986
987
988
@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,
    )

989
    processor = MULTIMODAL_REGISTRY.create_processor(model_config)
990
991

    rng = np.random.RandomState(0)
992
    image = random_image(rng, min_wh=128, max_wh=256)
993
994
995
996
997
998
999
1000
1001
1002
    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:
1003
        exc_ctx = pytest.raises(ValueError, match="At most")
1004
1005
1006
1007
1008
1009
1010

    with exc_ctx:
        processor.apply(
            "<image>" * num_images,
            mm_data=mm_data,
            hf_processor_mm_kwargs={},
        )
1011
1012


1013
class DummyProcessor:
1014

1015
    def __init__(self, a: int = 0, b: int = 0) -> None:
1016
1017
        super().__init__()

1018
1019
        self.a = a
        self.b = b
1020
1021
1022

    def __call__(
        self,
1023
1024
1025
1026
1027
        a: int = 0,
        c: int = 0,
        return_tensors: Optional[str] = None,
    ) -> dict[str, int]:
        return dict(a=a, c=c)
1028
1029
1030


# yapf: disable
1031
@pytest.mark.parametrize("model_id", ["Qwen/Qwen2-VL-2B-Instruct"])  # Dummy
1032
@pytest.mark.parametrize(
1033
    ("config_kwargs", "inference_kwargs", "expected_kwargs"),
1034
    [
1035
1036
1037
1038
1039
1040
1041
        ({"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}),
1042
1043
1044
    ],
)
# yapf: enable
1045
1046
1047
1048
1049
1050
1051
1052
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())
1053

1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
    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
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
# 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())
1091

1092
1093
1094
    ctx = InputProcessingContext(
        model_config=ModelConfig(model_id, mm_processor_kwargs=config_kwargs),
        tokenizer=mock_tokenizer,
1095
1096
    )

1097
1098
1099
1100
    processor = ctx.get_hf_processor(DummyProcessor)  # type: ignore[arg-type]

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