inputs.py 24.9 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
from abc import ABC, abstractmethod
4
from collections import UserDict, defaultdict
5
6
from collections.abc import Mapping, Sequence
from dataclasses import dataclass
7
from functools import partial
8
from itertools import accumulate
9
10
from typing import (TYPE_CHECKING, Any, Literal, Optional, TypedDict, TypeVar,
                    Union, cast, final)
11
12
13
14
15

import numpy as np
import torch
import torch.types
from PIL.Image import Image
16
from transformers import BatchFeature
17
from typing_extensions import NotRequired, TypeAlias
18

19
20
from vllm.jsontree import JSONTree, json_map_leaves
from vllm.utils import full_groupby, is_list_of
21

22
23
24
if TYPE_CHECKING:
    from .hasher import MultiModalHashDict

25
26
_T = TypeVar("_T")

27
HfImageItem: TypeAlias = Union[Image, np.ndarray, torch.Tensor]
28
"""
29
30
A :class:`transformers.image_utils.ImageInput` representing a single image
item, which can be passed to a HuggingFace :code:`ImageProcessor`.
31
32
"""

33
34
HfVideoItem: TypeAlias = Union[list[Image], np.ndarray, torch.Tensor,
                               list[np.ndarray], list[torch.Tensor]]
35
"""
36
37
A :class:`transformers.image_utils.VideoInput` representing a single video
item, which can be passed to a HuggingFace :code:`VideoProcessor`.
38
39
"""

40
HfAudioItem: TypeAlias = Union[list[float], np.ndarray, torch.Tensor]
41
"""
42
43
Represents a single audio
item, which can be passed to a HuggingFace :code:`AudioProcessor`.
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
ImageItem: TypeAlias = Union[HfImageItem, torch.Tensor]
"""
A :class:`transformers.image_utils.ImageInput` representing a single image
item, which can be passed to a HuggingFace :code:`ImageProcessor`.

Alternatively, a 3-D tensor or batch of 2-D tensors,
which are treated as image embeddings;
these are directly passed to the model without HF processing.
"""

VideoItem: TypeAlias = Union[HfVideoItem, torch.Tensor]
"""
A :class:`transformers.image_utils.VideoInput` representing a single video
item, which can be passed to a HuggingFace :code:`VideoProcessor`.

Alternatively, a 3-D tensor or batch of 2-D tensors,
which are treated as video embeddings;
these are directly passed to the model without HF processing.
"""

AudioItem: TypeAlias = Union[HfAudioItem, tuple[np.ndarray, float],
                             torch.Tensor]
"""
Represents a single audio
item, which can be passed to a HuggingFace :code:`AudioProcessor`.

Alternatively, a tuple `(audio, sampling_rate)`, where the sampling rate
is different from that expected by the model;
these are resampled to the model's sampling rate before being processed by HF.

Alternatively, a 3-D tensor or batch of 2-D tensors,
which are treated as audio embeddings;
these are directly passed to the model without HF processing.
"""

ModalityData: TypeAlias = Union[_T, list[_T]]
82
83
84
85
86
87
88
89
90
91
92
93
"""
Either a single data item, or a list of data items.

The number of data items allowed per modality is restricted by
:code:`--limit-mm-per-prompt`.
"""


@final
class MultiModalDataBuiltins(TypedDict, total=False):
    """Type annotations for modality types predefined by vLLM."""

94
    image: ModalityData[ImageItem]
95
96
    """The input image(s)."""

97
    video: ModalityData[VideoItem]
98
99
    """The input video(s)."""

100
    audio: ModalityData[AudioItem]
101
102
103
    """The input audio(s)."""


104
MultiModalDataDict: TypeAlias = Mapping[str, ModalityData[Any]]
105
106
"""
A dictionary containing an entry for each modality type to input.
107
108

The built-in modalities are defined by :class:`MultiModalDataBuiltins`.
109
110
111
"""


112
113
@dataclass(frozen=True)
class PlaceholderRange:
114
115
116
    """
    Placeholder location information for multi-modal data.

117
118
119
120
    Example:

        Prompt: :code:`AAAA BBBB What is in these images?`

121
        Images A and B will have:
122
123
124

        .. code-block::

125
126
            A: PlaceholderRange(offset=0, length=4)
            B: PlaceholderRange(offset=5, length=4)
127
128
129
130
131
132
133
134
    """

    offset: int
    """The start index of the placeholder in the prompt."""

    length: int
    """The length of the placeholder."""

135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
    is_embed: Optional[torch.Tensor] = None
    """
    A boolean mask of shape `(length,)` indicating which positions
    between `offset` and `offset + length` to assign embeddings to.
    """

    def get_num_embeds(self) -> int:
        if self.is_embed is None:
            return self.length

        return int(self.is_embed.sum().item())

    def __eq__(self, other: object) -> bool:
        if not isinstance(other, self.__class__):
            return False
        if not (self.offset, self.length) == (other.offset, other.length):
            return False

        if self.is_embed is None:
            return other.is_embed is None
        if other.is_embed is None:
            return self.is_embed is None

        return nested_tensors_equal(self.is_embed, other.is_embed)

160

161
162
NestedTensors = Union[list["NestedTensors"], list[torch.Tensor], torch.Tensor,
                      tuple[torch.Tensor, ...]]
163
164
165
166
"""
Uses a list instead of a tensor if the dimensions of each element do not match.
"""

167
168
169
170

def nested_tensors_equal(a: NestedTensors, b: NestedTensors) -> bool:
    """Equality check between :data:`NestedTensors` objects."""
    if isinstance(a, torch.Tensor):
171
        return isinstance(b, torch.Tensor) and torch.equal(a, b)
172
    elif isinstance(b, torch.Tensor):
173
        return isinstance(a, torch.Tensor) and torch.equal(b, a)
174
175
176
177
178
179
180
181
182
183
184
185
186

    if isinstance(a, list):
        return (isinstance(b, list)
                and all(nested_tensors_equal(a_, b_) for a_, b_ in zip(a, b)))
    if isinstance(b, list):
        return (isinstance(a, list)
                and all(nested_tensors_equal(b_, a_) for b_, a_ in zip(b, a)))

    # Both a and b are scalars
    return a == b


BatchedTensorInputs: TypeAlias = Mapping[str, NestedTensors]
187
188
189
190
191
192
"""
A dictionary containing nested tensors which have been batched via
:meth:`MultiModalKwargs.batch`.
"""


193
@dataclass(frozen=True)
194
class MultiModalFieldElem:
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
    """
    Represents a keyword argument corresponding to a multi-modal item
    in :class:`MultiModalKwargs`.
    """

    modality: str
    """
    The modality of the corresponding multi-modal item.
    Each multi-modal item can consist of multiple keyword arguments.
    """

    key: str
    """
    The key of this field in :class:`MultiModalKwargs`,
    i.e. the name of the keyword argument to be passed to the model.
    """

212
    data: NestedTensors
213
214
215
216
217
218
219
220
221
222
    """
    The tensor data of this field in :class:`MultiModalKwargs`,
    i.e. the value of the keyword argument to be passed to the model.
    """

    field: "BaseMultiModalField"
    """
    Defines how to combine the tensor data of this field with others
    in order to batch multi-modal items together for model inference.
    """
223
224
225
226
227

    def __eq__(self, other: object) -> bool:
        if not isinstance(other, self.__class__):
            return False

228
229
230
        return ((self.modality, self.key) == (other.modality, other.key)
                and nested_tensors_equal(self.data, other.data)
                and type(self.field) == type(other.field))  # noqa: E721
231
232
233
234


@dataclass(frozen=True)
class BaseMultiModalField(ABC):
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
    """
    Defines how to interpret tensor data belonging to a keyword argument in
    :class:`MultiModalKwargs` for multiple multi-modal items, and vice versa.
    """

    def _field_factory(self, *, modality: str, key: str):
        f = partial(
            MultiModalFieldElem,
            modality=modality,
            key=key,
            field=self,
        )

        # Allow passing data as positional argument
        def factory(data: NestedTensors) -> MultiModalFieldElem:
            return f(data=data)

        return factory
253
254

    @abstractmethod
255
256
257
258
259
260
261
262
263
264
265
266
    def build_elems(
        self,
        modality: str,
        key: str,
        data: NestedTensors,
    ) -> Sequence[MultiModalFieldElem]:
        """
        Construct :class:`MultiModalFieldElem` instances to represent
        the provided data.
        
        This is the inverse of :meth:`reduce_data`.
        """
267
268
        raise NotImplementedError

269
270
271
    @abstractmethod
    def _reduce_data(self, batch: list[NestedTensors]) -> NestedTensors:
        raise NotImplementedError
272

273
274
275
    def reduce_data(self, elems: list[MultiModalFieldElem]) -> NestedTensors:
        """
        Merge the data from multiple instances of :class:`MultiModalFieldElem`.
276

277
278
279
280
281
        This is the inverse of :meth:`build_elems`.
        """
        field_types = [type(item.field) for item in elems]
        if len(set(field_types)) > 1:
            raise ValueError(f"Cannot merge different {field_types=}")
282

283
        return self._reduce_data([item.data for item in elems])
284
285
286
287
288


@dataclass(frozen=True)
class MultiModalBatchedField(BaseMultiModalField):
    """
289
290
    See also:
        :func:`MultiModalFieldConfig.batched`
291
292
    """

293
294
295
296
297
298
299
300
    def build_elems(
        self,
        modality: str,
        key: str,
        data: NestedTensors,
    ) -> Sequence[MultiModalFieldElem]:
        field_factory = self._field_factory(modality=modality, key=key)
        return [field_factory(item) for item in data]
301
302
303

    def _reduce_data(self, batch: list[NestedTensors]) -> NestedTensors:
        if len(batch) > 0 and is_list_of(batch, torch.Tensor, check="all"):
304
305
306
307
308
            if len(batch) == 1:
                # An optimization when `batch` contains only one tensor:
                # - produce exactly same result as `torch.stack(batch)`
                # - will achieve zero-copy if the tensor is contiguous
                return batch[0].unsqueeze(0).contiguous()
309
            first_shape = batch[0].shape
310
            if all(elem.shape == first_shape for elem in batch):
311
312
313
314
315
316
317
318
                return torch.stack(batch)

        return batch


@dataclass(frozen=True)
class MultiModalFlatField(BaseMultiModalField):
    """
319
320
321
    See also:
        :func:`MultiModalFieldConfig.flat`
        :func:`MultiModalFieldConfig.flat_from_sizes`
322
    """
323
    slices: Sequence[slice]
324

325
    def build_elems(
326
        self,
327
328
329
330
331
332
        modality: str,
        key: str,
        data: NestedTensors,
    ) -> Sequence[MultiModalFieldElem]:
        field_factory = self._field_factory(modality=modality, key=key)
        return [field_factory(data[s]) for s in self.slices]
333
334
335

    def _reduce_data(self, batch: list[NestedTensors]) -> NestedTensors:
        if len(batch) > 0 and is_list_of(batch, torch.Tensor, check="all"):
336
337
338
339
340
            if len(batch) == 1:
                # An optimization when `batch` contains only one tensor:
                # - produce exactly same result as `torch.concat(batch)`
                # - will achieve zero-copy if the tensor is contiguous
                return batch[0].contiguous()
341
            first_shape = batch[0].shape
342
            if all(elem.shape[1:] == first_shape[1:] for elem in batch):
343
344
                return torch.concat(batch)

345
        return [e for elem in batch for e in elem]
346
347


348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
@dataclass(frozen=True)
class MultiModalSharedField(BaseMultiModalField):
    """
    See also:
        :func:`MultiModalFieldConfig.shared`
    """
    batch_size: int

    def build_elems(
        self,
        modality: str,
        key: str,
        data: NestedTensors,
    ) -> Sequence[MultiModalFieldElem]:
        field_factory = self._field_factory(modality=modality, key=key)
        return [field_factory(data)] * self.batch_size

    def _reduce_data(self, batch: list[NestedTensors]) -> NestedTensors:
        return batch[0]


369
370
371
372
class MultiModalFieldConfig:

    @staticmethod
    def batched(modality: str):
373
374
375
376
377
378
379
380
381
382
        """
        Defines a field where an element in the batch is obtained by
        indexing into the first dimension of the underlying data.

        Args:
            modality: The modality of the multi-modal item that uses this
                keyword argument.

        Example:

383
            .. code-block::
384

385
386
387
388
                Input:
                    Data: [[AAAA]
                        [BBBB]
                        [CCCC]]
389

390
391
392
393
                Output:
                    Element 1: [AAAA]
                    Element 2: [BBBB]
                    Element 3: [CCCC]
394
        """
395
        return MultiModalFieldConfig(
396
            field=MultiModalBatchedField(),
397
398
399
400
401
            modality=modality,
        )

    @staticmethod
    def flat(modality: str, slices: Sequence[slice]):
402
403
404
405
406
407
408
409
410
411
412
413
        """
        Defines a field where an element in the batch is obtained by
        slicing along the first dimension of the underlying data.

        Args:
            modality: The modality of the multi-modal item that uses this
                keyword argument.
            slices: For each multi-modal item, a slice that is used to extract
                the data corresponding to it.

        Example:

414
415
416
417
            .. code-block::
        
                Given:
                    slices: [slice(0, 3), slice(3, 7), slice(7, 9)]
418

419
420
                Input:
                    Data: [AAABBBBCC]
421

422
423
424
425
                Output:
                    Element 1: [AAA]
                    Element 2: [BBBB]
                    Element 3: [CC]
426
        """
427
        return MultiModalFieldConfig(
428
            field=MultiModalFlatField(slices=slices),
429
430
431
            modality=modality,
        )

432
433
    @staticmethod
    def flat_from_sizes(modality: str, size_per_item: torch.Tensor):
434
435
436
437
438
439
440
441
442
443
444
445
        """
        Defines a field where an element in the batch is obtained by
        slicing along the first dimension of the underlying data.

        Args:
            modality: The modality of the multi-modal item that uses this
                keyword argument.
            slices: For each multi-modal item, the size of the slice that
                is used to extract the data corresponding to it.

        Example:

446
447
448
449
            .. code-block::
        
                Given:
                    size_per_item: [3, 4, 2]
450

451
452
                Input:
                    Data: [AAABBBBCC]
453

454
455
456
457
                Output:
                    Element 1: [AAA]
                    Element 2: [BBBB]
                    Element 3: [CC]
458
459
460
461
462
    
        See also:
            :func:`MultiModalFieldConfig.flat`
        """

463
464
465
466
        if size_per_item.ndim != 1:
            raise ValueError("size_per_item should be a 1-D tensor, "
                             f"but found shape: {size_per_item.shape}")

467
468
469
470
471
472
473
474
        slice_idxs = [0, *accumulate(size_per_item)]
        slices = [
            slice(slice_idxs[i], slice_idxs[i + 1])
            for i in range(len(size_per_item))
        ]

        return MultiModalFieldConfig.flat(modality, slices)

475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
    @staticmethod
    def shared(modality: str, batch_size: int):
        """
        Defines a field where an element in the batch is obtained by
        taking the entirety of the underlying data.

        This means that the data is the same for each element in the batch.

        Args:
            modality: The modality of the multi-modal item that uses this
                keyword argument.
            batch_size: The number of multi-modal items which share this data.

        Example:

490
491
492
493
            .. code-block::
        
                Given:
                    batch_size: 4
494

495
496
                Input:
                    Data: [XYZ]
497

498
499
500
501
502
                Output:
                    Element 1: [XYZ]
                    Element 2: [XYZ]
                    Element 3: [XYZ]
                    Element 4: [XYZ]
503
504
505
506
507
508
509
        """
        return MultiModalFieldConfig(
            field=MultiModalSharedField(batch_size),
            modality=modality,
        )

    def __init__(self, field: BaseMultiModalField, modality: str) -> None:
510
511
        super().__init__()

512
        self.field = field
513
        self.modality = modality
514

515
    def build_elems(
516
517
518
        self,
        key: str,
        batch: NestedTensors,
519
    ) -> Sequence[MultiModalFieldElem]:
520
        return self.field.build_elems(self.modality, key, batch)
521
522


523
524
525
526
class MultiModalKwargsItem(UserDict[str, MultiModalFieldElem]):
    """
    A collection of :class:`MultiModalFieldElem`
    corresponding to a data item in :class:`MultiModalDataItems`.
527
    """
528

529
530
    @staticmethod
    def from_elems(elems: Sequence[MultiModalFieldElem]):
531
        return MultiModalKwargsItem({elem.key: elem for elem in elems})
532

533
534
    @property
    def modality(self) -> str:
535
        modalities = {elem.modality for elem in self.data.values()}
536
537
        assert len(modalities) == 1, f"Found different modalities={modalities}"
        return next(iter(modalities))
538
539


540
541
542
543
544
545
# NOTE: UserDict is for V0 compatibility.
# V1 should access individual items via `get_item`.
class MultiModalKwargs(UserDict[str, NestedTensors]):
    """
    A dictionary that represents the keyword arguments to
    :meth:`~torch.nn.Module.forward`.
546

547
548
549
    The metadata :code:`items` enables us to obtain the keyword arguments
    corresponding to each data item in :class:`MultiModalDataItems`, via
    :meth:`get_item` and :meth:`get_items`.
550
551
    """

552
553
554
555
556
557
558
    @staticmethod
    def from_hf_inputs(
        hf_inputs: BatchFeature,
        config_by_key: Mapping[str, MultiModalFieldConfig],
    ):
        # NOTE: This skips fields in `hf_inputs` that are not in `config_by_key`
        # We assume that those fields are not used in vLLM
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
        elems_by_key = dict[str, Sequence[MultiModalFieldElem]]()
        keys_by_modality = defaultdict[str, set[str]](set)
        for key, config in config_by_key.items():
            batch = hf_inputs.get(key)
            if batch is not None:
                elems = config.build_elems(key, batch)
                if len(elems) > 0:
                    elems_by_key[key] = elems
                    keys_by_modality[config.modality].add(key)

        items = list[MultiModalKwargsItem]()
        for modality, keys in keys_by_modality.items():
            elems_in_modality = {k: elems_by_key[k] for k in keys}
            batch_sizes = {k: len(v) for k, v in elems_in_modality.items()}

            if len(set(batch_sizes.values())) > 1:
                raise ValueError(
                    f"Cannot merge different batch sizes for {modality=}! "
                    f"Found: {batch_sizes=}")

            batch_size = next(iter(batch_sizes.values()))
            for item_idx in range(batch_size):
                elems = [v[item_idx] for v in elems_in_modality.values()]
                items.append(MultiModalKwargsItem.from_elems(elems))

        return MultiModalKwargs.from_items(items)
585
586

    @staticmethod
587
588
589
590
591
592
593
    def from_items(items: Sequence[MultiModalKwargsItem]):
        """Construct a new :class:`MultiModalKwargs` from multiple items."""
        elems_by_key = defaultdict[str, list[MultiModalFieldElem]](list)
        for item in items:
            for key, elem in item.items():
                elems_by_key[key].append(elem)

594
        data = {
595
            key: elems[0].field.reduce_data(elems)
596
            for key, elems in elems_by_key.items() if len(elems) > 0
597
598
        }

599
        return MultiModalKwargs(data, items=items)
600
601
602
603
604

    def __init__(
        self,
        data: Mapping[str, NestedTensors],
        *,
605
        items: Optional[Sequence[MultiModalKwargsItem]] = None,
606
607
608
    ) -> None:
        super().__init__(data)

609
610
        items_by_modality = full_groupby(items or [], key=lambda x: x.modality)
        self._items_by_modality = dict(items_by_modality)
611

612
613
614
    @property
    def modalities(self):
        return self._items_by_modality.keys()
615

616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
    @staticmethod
    def _try_stack(nested_tensors: NestedTensors) -> NestedTensors:
        """
        Stack the inner dimensions that have the same shape in
        a nested list of tensors.

        Thus, a dimension represented by a list means that the inner
        dimensions are different for each element along that dimension.
        """
        if isinstance(nested_tensors, torch.Tensor):
            return nested_tensors

        # TODO: Remove these once all models have been migrated
        if isinstance(nested_tensors, np.ndarray):
            return torch.from_numpy(nested_tensors)
        if isinstance(nested_tensors, (int, float)):
            return torch.tensor(nested_tensors)

        stacked = [MultiModalKwargs._try_stack(t) for t in nested_tensors]
        if not is_list_of(stacked, torch.Tensor, check="all"):
            # Only tensors (not lists) can be stacked.
            return stacked

639
        tensors_ = cast(list[torch.Tensor], stacked)
640
641
642
643
644
645
        if len(tensors_) == 1:
            # An optimization when `tensors_` contains only one tensor:
            # - produce exactly same result as `torch.stack(tensors_)`
            # - will achieve zero-copy if the tensor is contiguous
            return tensors_[0].unsqueeze(0).contiguous()

646
647
648
649
650
651
652
        if any(t.shape != tensors_[0].shape for t in tensors_):
            # The tensors have incompatible shapes and can't be stacked.
            return tensors_

        return torch.stack(tensors_)

    @staticmethod
653
    def batch(inputs_list: list["MultiModalKwargs"]) -> BatchedTensorInputs:
654
655
656
657
658
659
660
661
662
663
664
665
666
667
        """
        Batch multiple inputs together into a dictionary.

        The resulting dictionary has the same keys as the inputs.
        If the corresponding value from each input is a tensor and they all
        share the same shape, the output value is a single batched tensor;
        otherwise, the output value is a list containing the original value
        from each input.
        """
        if len(inputs_list) == 0:
            return {}

        # We need to consider the case where each item in the batch
        # contains different modalities (i.e. different keys).
668
        item_lists = defaultdict[str, list[NestedTensors]](list)
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

        for inputs in inputs_list:
            for k, v in inputs.items():
                item_lists[k].append(v)

        return {
            k: MultiModalKwargs._try_stack(item_list)
            for k, item_list in item_lists.items()
        }

    @staticmethod
    def as_kwargs(
        batched_inputs: BatchedTensorInputs,
        *,
        device: torch.types.Device,
    ) -> BatchedTensorInputs:
        json_inputs = cast(JSONTree[torch.Tensor], batched_inputs)

        json_mapped = json_map_leaves(
            lambda x: x.to(device, non_blocking=True),
            json_inputs,
        )

        return cast(BatchedTensorInputs, json_mapped)

694
695
696
697
698
699
700
    def __delitem__(self, key: str) -> None:
        super().__delitem__(key)

        for items in self._items_by_modality.values():
            for item in items:
                item.pop(key, None)

701
702
703
    def __eq__(self, other: object) -> bool:
        if not isinstance(other, self.__class__):
            return False
704
        if self._items_by_modality != other._items_by_modality:
705
706
707
708
709
710
            return False

        ks = self.keys()
        return (ks == other.keys()
                and all(nested_tensors_equal(self[k], other[k]) for k in ks))

711
712
713
714
715
    def _validate_modality(self, method_name: str, modality: str) -> None:
        if not self._items_by_modality:
            raise RuntimeError(
                f"`{method_name}` is not supported when "
                "MultiModalKwargs is not initialized with `items`")
716

717
718
        if modality not in self._items_by_modality:
            available_modalities = set(self._items_by_modality.keys())
719
720
721
            raise KeyError(f"Modality {modality!r} not found. "
                           f"Available modalities: {available_modalities}")

722
723
724
725
    def get_item_count(self, modality: str) -> int:
        """Get the number of items belonging to a modality."""
        self._validate_modality("get_item_count", modality)
        return len(self._items_by_modality[modality])
726

727
728
729
730
731
732
733
    def get_item(self, modality: str, item_index: int) -> MultiModalKwargsItem:
        """
        Get the keyword arguments corresponding to an item identified by
        its modality and index.
        """
        self._validate_modality("get_item", modality)
        return self._items_by_modality[modality][item_index]
734

735
    def get_items(self, modality: str) -> Sequence[MultiModalKwargsItem]:
736
        """
737
738
        Get the keyword arguments corresponding to each item belonging to
        a modality.
739
        """
740
741
        self._validate_modality("get_items", modality)
        return self._items_by_modality[modality]
742

743
744
745

MultiModalPlaceholderDict = Mapping[str, Sequence[PlaceholderRange]]
"""
746
A dictionary containing placeholder ranges for each modality.
747
748
749
"""


750
class MultiModalInputs(TypedDict):
751
    """
752
753
    Represents the outputs of
    :class:`vllm.multimodal.processing.BaseMultiModalProcessor`,
754
755
756
757
758
759
760
    ready to be passed to vLLM internals.
    """

    type: Literal["multimodal"]
    """The type of inputs."""

    prompt: str
761
    """The processed prompt text."""
762

763
    prompt_token_ids: list[int]
764
765
    """The processed token IDs which includes placeholder tokens."""

766
    token_type_ids: NotRequired[list[int]]
767
768
    """The token type IDs of the prompt."""

769
770
771
    mm_kwargs: MultiModalKwargs
    """Keyword arguments to be directly passed to the model after batching."""

772
    mm_hashes: Optional["MultiModalHashDict"]
773
774
    """The hashes of the multi-modal data."""

775
776
777
778
779
    mm_placeholders: MultiModalPlaceholderDict
    """
    For each modality, information about the placeholder tokens in
    :code:`prompt_token_ids`.
    """
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795


class MultiModalEncDecInputs(MultiModalInputs):
    """
    Represents the outputs of :class:`vllm.multimodal.EncDecMultiModalProcessor`
    ready to be passed to vLLM internals.
    """

    encoder_prompt: str
    """The processed encoder prompt text."""

    encoder_prompt_token_ids: list[int]
    """The processed token IDs of the encoder prompt."""

    encoder_token_type_ids: NotRequired[list[int]]
    """The token type IDs of the encoder prompt."""