interfaces.py 30.9 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
from collections.abc import Callable, Iterable, Mapping, MutableSequence, Set
5
6
7
8
9
from typing import (
    TYPE_CHECKING,
    ClassVar,
    Literal,
    Protocol,
10
    TypeAlias,
11
12
13
    overload,
    runtime_checkable,
)
14

15
import numpy as np
16
import torch
17
import torch.nn as nn
18
from torch import Tensor
19
from transformers.models.whisper.tokenization_whisper import LANGUAGES
20
from typing_extensions import Self, TypeIs
21

22
from vllm.config import ModelConfig, SpeechToTextConfig
23
from vllm.inputs import TokensPrompt
24
from vllm.inputs.data import PromptType
25
from vllm.logger import init_logger
26
from vllm.model_executor.layers.quantization import QuantizationConfig
27
from vllm.utils.func_utils import supports_kw
28

29
from .interfaces_base import VllmModel, is_pooling_model
30

31
if TYPE_CHECKING:
32
    from vllm.config import VllmConfig
33
    from vllm.model_executor.models.utils import WeightsMapper
34
    from vllm.multimodal.inputs import MultiModalFeatureSpec
35
    from vllm.sequence import IntermediateTensors
36
37
38
else:
    VllmConfig = object
    WeightsMapper = object
39
    MultiModalFeatureSpec = object
40
    IntermediateTensors = object
41

42
43
logger = init_logger(__name__)

44
MultiModalEmbeddings: TypeAlias = list[Tensor] | Tensor | tuple[Tensor, ...]
45
46
47
48
49
50
51
"""
The output embeddings must be one of the following formats:

- A list or tuple of 2D tensors, where each tensor corresponds to
    each input multimodal data item (e.g, image).
- A single 3D tensor, with the batch dimension grouping the 2D tensors.
"""
52

53
54

@runtime_checkable
55
class SupportsMultiModal(Protocol):
56
    """The interface required for all multi-modal models."""
57

58
    supports_multimodal: ClassVar[Literal[True]] = True
59
    """
60
    A flag that indicates this model supports multi-modal inputs.
61
62
63
64
65

    Note:
        There is no need to redefine this flag if this class is in the
        MRO of your model class.
    """
66

67
68
69
70
71
72
    supports_multimodal_raw_input_only: ClassVar[bool] = False
    """
    A flag that indicates this model supports multi-modal inputs and processes
    them in their raw form and not embeddings.
    """

73
74
75
76
77
78
    supports_encoder_tp_data: ClassVar[bool] = False
    """
    A flag that indicates whether this model supports
    `multimodal_config.mm_encoder_tp_mode="data"`.
    """

79
80
81
82
83
84
    merge_by_field_config: ClassVar[bool] = False
    """
    A flag that indicates which implementation of
    `vllm.multimodal.utils.group_mm_kwargs_by_modality` to use.
    """

85
86
87
88
89
    multimodal_cpu_fields: ClassVar[Set[str]] = frozenset()
    """
    A set indicating CPU-only multimodal fields.
    """

90
    @classmethod
91
    def get_placeholder_str(cls, modality: str, i: int) -> str | None:
92
93
94
95
96
        """
        Get the placeholder text for the `i`th `modality` item in the prompt.
        """
        ...

97
    def get_multimodal_embeddings(self, **kwargs: object) -> MultiModalEmbeddings:
98
        """
99
        Returns multimodal embeddings generated from multimodal kwargs
100
        to be merged with text embeddings.
101

102
        Note:
103
104
            The returned multimodal embeddings must be in the same order as
            the appearances of their corresponding multimodal data item in the
105
            input prompt.
106
107
108
        """
        ...

109
    def get_language_model(self) -> VllmModel:
110
111
112
        """
        Returns the underlying language model used for text generation.

113
        This is typically the `torch.nn.Module` instance responsible for
114
115
116
117
118
119
120
        processing the merged multimodal embeddings and producing hidden states

        Returns:
            torch.nn.Module: The core language model component.
        """
        ...

121
    @overload
122
    def get_input_embeddings(self, input_ids: Tensor) -> Tensor: ...
123
124
125
126
127
128
129
130
131

    @overload
    def get_input_embeddings(
        self,
        input_ids: Tensor,
        multimodal_embeddings: MultiModalEmbeddings,
        *,
        is_multimodal: torch.Tensor,
        handle_oov_mm_token: bool = False,
132
    ) -> Tensor: ...
133
134
135
136
137
138

    def _get_text_embeddings(
        self,
        input_ids: Tensor,
        get_input_embeddings: Callable[[Tensor], Tensor],
        *,
139
        is_multimodal: Tensor | None,
140
141
142
143
144
145
146
147
148
149
150
151
152
153
        handle_oov_mm_token: bool,
    ) -> Tensor:
        if handle_oov_mm_token and is_multimodal is not None:
            is_text = ~is_multimodal
            text_embeds = get_input_embeddings(input_ids[is_text])

            return torch.empty(
                (input_ids.shape[0], text_embeds.shape[1]),
                dtype=text_embeds.dtype,
                device=text_embeds.device,
            ).masked_scatter_(is_text.unsqueeze_(-1), text_embeds)

        return get_input_embeddings(input_ids)

154
155
    def get_input_embeddings(
        self,
156
        input_ids: Tensor,
157
        multimodal_embeddings: MultiModalEmbeddings | None = None,
158
        *,
159
        is_multimodal: Tensor | None = None,
160
        handle_oov_mm_token: bool = False,
161
    ) -> Tensor:
162
        """
163
164
165
166
167
168
169
170
171
172
        Apply token embeddings to `input_ids`.

        If `multimodal_embeddings` is passed, scatter them into
        `input_ids` according to the mask `is_multimodal`.

        In case the multi-modal token IDs exceed the vocabulary size of
        the language model, you can set `handle_oov_mm_token=False`
        to avoid calling the language model's `get_input_embeddings` method
        on those tokens. Note however that doing so increases memory usage
        as an additional buffer is needed to hold the input embeddings.
173
        """
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
        from .utils import _merge_multimodal_embeddings

        inputs_embeds = self._get_text_embeddings(
            input_ids,
            self.get_language_model().get_input_embeddings,
            is_multimodal=is_multimodal,
            handle_oov_mm_token=handle_oov_mm_token,
        )

        if multimodal_embeddings is None or len(multimodal_embeddings) == 0:
            return inputs_embeds

        if is_multimodal is None:
            raise ValueError(
                "`get_input_embeddings` now requires `is_multimodal` arg, "
                "please update your model runner according to "
190
191
                "https://github.com/vllm-project/vllm/pull/16229."
            )
192
193
194
195
196
197

        return _merge_multimodal_embeddings(
            inputs_embeds=inputs_embeds,
            multimodal_embeddings=multimodal_embeddings,
            is_multimodal=is_multimodal,
        )
198

199

200
201
202
203
204
205
@runtime_checkable
class SupportsMultiModalPruning(Protocol):
    """The interface required for models that support returning both input
    embeddings and positions. Model may require custom positions for dynamic
    pruning of multimodal embeddings.
    """
206

207
208
209
    supports_multimodal_pruning: ClassVar[Literal[True]] = True

    def recompute_mrope_positions(
210
211
212
213
214
        self,
        input_ids: list[int],
        multimodal_embeddings: MultiModalEmbeddings,
        mrope_positions: torch.LongTensor,
        num_computed_tokens: int,
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
    ) -> tuple[MultiModalEmbeddings, Tensor, int]:
        """
        Update part of input mrope positions (starting with
        num_computed_tokens index). Original mrope_positions are computed
        for unpruned sequence and becomes incorrect once pruning occurs,
        so once we prune media tokens we should reflect this in the
        mrope_positions before we feed it to LLM.

        Args:
            input_ids: (N,) All input tokens of the prompt containing
                entire sequence.
            multimodal_embeddings: Tuple of multimodal embeddings that
                fits into the prefill chunk that is being processed.
            mrope_positions: Existing mrope positions (3, N) for entire
                sequence
            num_computed_tokens: A number of computed tokens so far.

        Returns:
            Tuple of (multimodal_embeddings, mrope_positions,
                mrope_position_delta).
        """
        ...


239
@overload
240
def supports_multimodal(model: type[object]) -> TypeIs[type[SupportsMultiModal]]: ...
241
242
243


@overload
244
def supports_multimodal(model: object) -> TypeIs[SupportsMultiModal]: ...
245
246


247
def supports_multimodal(
248
249
    model: type[object] | object,
) -> TypeIs[type[SupportsMultiModal]] | TypeIs[SupportsMultiModal]:
250
    return getattr(model, "supports_multimodal", False)
251
252


253
def supports_multimodal_raw_input_only(model: type[object] | object) -> bool:
254
    return getattr(model, "supports_multimodal_raw_input_only", False)
255

256

257
def supports_multimodal_encoder_tp_data(model: type[object] | object) -> bool:
258
    return getattr(model, "supports_encoder_tp_data", False)
259
260


261
262
@overload
def supports_multimodal_pruning(
263
264
    model: type[object],
) -> TypeIs[type[SupportsMultiModalPruning]]: ...
265
266
267


@overload
268
def supports_multimodal_pruning(model: object) -> TypeIs[SupportsMultiModalPruning]: ...
269
270
271


def supports_multimodal_pruning(
272
273
    model: type[object] | object,
) -> TypeIs[type[SupportsMultiModalPruning]] | TypeIs[SupportsMultiModalPruning]:
274
275
276
    return getattr(model, "supports_multimodal_pruning", False)


277
278
279
280
281
282
283
284
285
286
287
288
289
290
@runtime_checkable
class SupportsScoreTemplate(Protocol):
    """The interface required for all models that support score template."""

    supports_score_template: ClassVar[Literal[True]] = True
    """
    A flag that indicates this model supports score template.

    Note:
        There is no need to redefine this flag if this class is in the
        MRO of your model class.
    """

    @classmethod
291
    def get_score_template(cls, query: str, document: str) -> str | None:
292
293
        """
        Generate a full prompt by populating the score template with query and document content.
294
        """  # noqa: E501
295
296
297
298
299
300
301
302
303
304
305
306
        ...

    @classmethod
    def post_process_tokens(cls, prompt: TokensPrompt) -> None:
        """
        Perform architecture-specific manipulations on the input tokens.
        """
        ...


@overload
def supports_score_template(
307
308
    model: type[object],
) -> TypeIs[type[SupportsScoreTemplate]]: ...
309
310
311


@overload
312
def supports_score_template(model: object) -> TypeIs[SupportsScoreTemplate]: ...
313
314
315


def supports_score_template(
316
317
    model: type[object] | object,
) -> TypeIs[type[SupportsScoreTemplate]] | TypeIs[SupportsScoreTemplate]:
318
    return getattr(model, "supports_score_template", False)
319
320


321
322
323
324
@runtime_checkable
class SupportsLoRA(Protocol):
    """The interface required for all models that support LoRA."""

325
326
327
328
329
330
331
332
    supports_lora: ClassVar[Literal[True]] = True
    """
    A flag that indicates this model supports LoRA.

    Note:
        There is no need to redefine this flag if this class is in the
        MRO of your model class.
    """
333
334
    # The `embedding_module` and `embedding_padding_modules`
    # are empty by default.
335
336
    embedding_modules: ClassVar[dict[str, str]] = {}
    embedding_padding_modules: ClassVar[list[str]] = []
337
    packed_modules_mapping: dict[str, list[str]] = {}
338
339
340
341
342
343
344
345


# We can't use runtime_checkable with ClassVar for issubclass checks
# so we need to treat the class as an instance and use isinstance instead
@runtime_checkable
class _SupportsLoRAType(Protocol):
    supports_lora: Literal[True]

346
347
348
    packed_modules_mapping: dict[str, list[str]]
    embedding_modules: dict[str, str]
    embedding_padding_modules: list[str]
349
350
351


@overload
352
def supports_lora(model: type[object]) -> TypeIs[type[SupportsLoRA]]: ...
353
354
355


@overload
356
def supports_lora(model: object) -> TypeIs[SupportsLoRA]: ...
357
358
359


def supports_lora(
360
361
    model: type[object] | object,
) -> TypeIs[type[SupportsLoRA]] | TypeIs[SupportsLoRA]:
362
363
364
365
366
367
368
369
    result = _supports_lora(model)

    if not result:
        lora_attrs = (
            "packed_modules_mapping",
            "embedding_modules",
            "embedding_padding_modules",
        )
370
        missing_attrs = tuple(attr for attr in lora_attrs if not hasattr(model, attr))
371
372
373
374
375
376
377
378
379
380
381
382
383

        if getattr(model, "supports_lora", False):
            if missing_attrs:
                logger.warning(
                    "The model (%s) sets `supports_lora=True`, "
                    "but is missing LoRA-specific attributes: %s",
                    model,
                    missing_attrs,
                )
        else:
            if not missing_attrs:
                logger.warning(
                    "The model (%s) contains all LoRA-specific attributes, "
384
385
386
                    "but does not set `supports_lora=True`.",
                    model,
                )
387
388
389
390

    return result


391
def _supports_lora(model: type[object] | object) -> bool:
392
393
394
395
    if isinstance(model, type):
        return isinstance(model, _SupportsLoRAType)

    return isinstance(model, SupportsLoRA)
396
397


398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
@runtime_checkable
class SupportsPP(Protocol):
    """The interface required for all models that support pipeline parallel."""

    supports_pp: ClassVar[Literal[True]] = True
    """
    A flag that indicates this model supports pipeline parallel.

    Note:
        There is no need to redefine this flag if this class is in the
        MRO of your model class.
    """

    def make_empty_intermediate_tensors(
        self,
        batch_size: int,
        dtype: torch.dtype,
        device: torch.device,
416
    ) -> IntermediateTensors:
417
418
419
420
421
        """Called when PP rank > 0 for profiling purposes."""
        ...

    def forward(
        self,
422
        *,
423
424
        intermediate_tensors: IntermediateTensors | None,
    ) -> IntermediateTensors | None:
425
        """
426
427
        Accept [`IntermediateTensors`][vllm.sequence.IntermediateTensors] when
        PP rank > 0.
428

429
430
        Return [`IntermediateTensors`][vllm.sequence.IntermediateTensors] only
        for the last PP rank.
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
        """
        ...


# We can't use runtime_checkable with ClassVar for issubclass checks
# so we need to treat the class as an instance and use isinstance instead
@runtime_checkable
class _SupportsPPType(Protocol):
    supports_pp: Literal[True]

    def make_empty_intermediate_tensors(
        self,
        batch_size: int,
        dtype: torch.dtype,
        device: torch.device,
446
    ) -> IntermediateTensors: ...
447
448
449

    def forward(
        self,
450
        *,
451
452
        intermediate_tensors: IntermediateTensors | None,
    ) -> Tensor | IntermediateTensors: ...
453
454
455


@overload
456
def supports_pp(model: type[object]) -> TypeIs[type[SupportsPP]]: ...
457
458
459


@overload
460
def supports_pp(model: object) -> TypeIs[SupportsPP]: ...
461
462
463


def supports_pp(
464
465
    model: type[object] | object,
) -> bool | TypeIs[type[SupportsPP]] | TypeIs[SupportsPP]:
466
467
468
469
470
471
    supports_attributes = _supports_pp_attributes(model)
    supports_inspect = _supports_pp_inspect(model)

    if supports_attributes and not supports_inspect:
        logger.warning(
            "The model (%s) sets `supports_pp=True`, but does not accept "
472
473
474
            "`intermediate_tensors` in its `forward` method",
            model,
        )
475
476

    if not supports_attributes:
477
478
        pp_attrs = ("make_empty_intermediate_tensors",)
        missing_attrs = tuple(attr for attr in pp_attrs if not hasattr(model, attr))
479
480
481
482
483
484
485
486
487
488
489
490
491

        if getattr(model, "supports_pp", False):
            if missing_attrs:
                logger.warning(
                    "The model (%s) sets `supports_pp=True`, "
                    "but is missing PP-specific attributes: %s",
                    model,
                    missing_attrs,
                )
        else:
            if not missing_attrs:
                logger.warning(
                    "The model (%s) contains all PP-specific attributes, "
492
493
494
                    "but does not set `supports_pp=True`.",
                    model,
                )
495
496
497
498

    return supports_attributes and supports_inspect


499
def _supports_pp_attributes(model: type[object] | object) -> bool:
500
501
502
503
504
505
    if isinstance(model, type):
        return isinstance(model, _SupportsPPType)

    return isinstance(model, SupportsPP)


506
def _supports_pp_inspect(model: type[object] | object) -> bool:
507
508
509
510
    model_forward = getattr(model, "forward", None)
    if not callable(model_forward):
        return False

511
    return supports_kw(model_forward, "intermediate_tensors")
512
513


514
515
516
517
518
519
520
521
@runtime_checkable
class HasInnerState(Protocol):
    """The interface required for all models that has inner state."""

    has_inner_state: ClassVar[Literal[True]] = True
    """
        A flag that indicates this model has inner state.
        Models that has inner state usually need access to the scheduler_config
522
        for max_num_seqs, etc. True for e.g. both Mamba and Jamba.
523
524
525
526
    """


@overload
527
def has_inner_state(model: object) -> TypeIs[HasInnerState]: ...
528
529
530


@overload
531
def has_inner_state(model: type[object]) -> TypeIs[type[HasInnerState]]: ...
532
533
534


def has_inner_state(
535
536
    model: type[object] | object,
) -> TypeIs[type[HasInnerState]] | TypeIs[HasInnerState]:
537
    return getattr(model, "has_inner_state", False)
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553


@runtime_checkable
class IsAttentionFree(Protocol):
    """The interface required for all models like Mamba that lack attention,
    but do have state whose size is constant wrt the number of tokens."""

    is_attention_free: ClassVar[Literal[True]] = True
    """
        A flag that indicates this model has no attention.
        Used for block manager and attention backend selection.
        True for Mamba but not Jamba.
    """


@overload
554
def is_attention_free(model: object) -> TypeIs[IsAttentionFree]: ...
555
556
557


@overload
558
def is_attention_free(model: type[object]) -> TypeIs[type[IsAttentionFree]]: ...
559
560
561


def is_attention_free(
562
563
    model: type[object] | object,
) -> TypeIs[type[IsAttentionFree]] | TypeIs[IsAttentionFree]:
564
    return getattr(model, "is_attention_free", False)
565
566


567
568
569
@runtime_checkable
class IsHybrid(Protocol):
    """The interface required for all models like Jamba that have both
570
    attention and mamba blocks, indicates that
571
572
573
574
575
576
577
578
    hf_config has 'layers_block_type'"""

    is_hybrid: ClassVar[Literal[True]] = True
    """
        A flag that indicates this model has both mamba and attention blocks
        , also indicates that the model's hf_config has 
        'layers_block_type' """

579
580
581
    @classmethod
    def get_mamba_state_shape_from_config(
        cls,
582
        vllm_config: VllmConfig,
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
        use_v1: bool = True,
    ) -> tuple[tuple[int, int], tuple[int, int, int]]:
        """Calculate shapes for Mamba's convolutional and state caches.

        Args:
            vllm_config: vLLM config
            use_v1: Get shapes for V1 (or V0)

        Returns:
            Tuple containing:
            - conv_state_shape: Shape for convolutional state cache
            - temporal_state_shape: Shape for state space model cache
        """
        ...

598
599

@overload
600
def is_hybrid(model: object) -> TypeIs[IsHybrid]: ...
601
602
603


@overload
604
def is_hybrid(model: type[object]) -> TypeIs[type[IsHybrid]]: ...
605
606
607


def is_hybrid(
608
609
    model: type[object] | object,
) -> TypeIs[type[IsHybrid]] | TypeIs[IsHybrid]:
610
    return getattr(model, "is_hybrid", False)
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
643
644
645
646
647
648
649
650
@runtime_checkable
class MixtureOfExperts(Protocol):
    """
    Check if the model is a mixture of experts (MoE) model.
    """

    expert_weights: MutableSequence[Iterable[Tensor]]
    """
    Expert weights saved in this rank.

    The first dimension is the layer, and the second dimension is different
    parameters in the layer, e.g. up/down projection weights.
    """

    num_moe_layers: int
    """Number of MoE layers in this model."""

    num_expert_groups: int
    """Number of expert groups in this model."""

    num_logical_experts: int
    """Number of logical experts in this model."""

    num_physical_experts: int
    """Number of physical experts in this model."""

    num_local_physical_experts: int
    """Number of local physical experts in this model."""

    num_routed_experts: int
    """Number of routed experts in this model."""

    num_shared_experts: int
    """Number of shared experts in this model."""

    num_redundant_experts: int
    """Number of redundant experts in this model."""

651
652
653
    moe_layers: Iterable[nn.Module]
    """List of MoE layers in this model."""

654
655
656
657
658
659
660
661
    def set_eplb_state(
        self,
        expert_load_view: Tensor,
        logical_to_physical_map: Tensor,
        logical_replica_count: Tensor,
    ) -> None:
        """
        Register the EPLB state in the MoE model.
662

663
664
665
666
667
668
669
670
671
672
673
674
675
        Since these are views of the actual EPLB state, any changes made by
        the EPLB algorithm are automatically reflected in the model's behavior
        without requiring additional method calls to set new states.

        You should also collect model's `expert_weights` here instead of in
        the weight loader, since after initial weight loading, further
        processing like quantization may be applied to the weights.

        Args:
            expert_load_view: A view of the expert load metrics tensor.
            logical_to_physical_map: Mapping from logical to physical experts.
            logical_replica_count: Count of replicas for each logical expert.
        """
676
677
678
679
680
681
682
683
684
        for layer_idx, layer in enumerate(self.moe_layers):
            # Register the expert weights.
            self.expert_weights.append(layer.get_expert_weights())
            layer.set_eplb_state(
                moe_layer_idx=layer_idx,
                expert_load_view=expert_load_view,
                logical_to_physical_map=logical_to_physical_map,
                logical_replica_count=logical_replica_count,
            )
685

686
687
688
689
    def update_physical_experts_metadata(
        self,
        num_physical_experts: int,
        num_local_physical_experts: int,
690
    ) -> None: ...
691

692
693

def is_mixture_of_experts(model: object) -> TypeIs[MixtureOfExperts]:
694
695
696
    return (
        isinstance(model, MixtureOfExperts) and getattr(model, "num_moe_layers", 0) > 0
    )
697
698


699
700
701
702
703
704
@runtime_checkable
class HasNoOps(Protocol):
    has_noops: ClassVar[Literal[True]] = True


@overload
705
def has_noops(model: object) -> TypeIs[HasNoOps]: ...
706
707
708


@overload
709
def has_noops(model: type[object]) -> TypeIs[type[HasNoOps]]: ...
710
711
712


def has_noops(
713
714
    model: type[object] | object,
) -> TypeIs[type[HasNoOps]] | TypeIs[HasNoOps]:
715
    return getattr(model, "has_noops", False)
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
745
@runtime_checkable
class SupportsMambaPrefixCaching(Protocol):
    """The interface for models whose mamba layers support prefix caching.

    This is currently experimental.
    """

    supports_mamba_prefix_caching: ClassVar[Literal[True]] = True


@overload
def supports_mamba_prefix_caching(
    model: object,
) -> TypeIs[SupportsMambaPrefixCaching]: ...


@overload
def supports_mamba_prefix_caching(
    model: type[object],
) -> TypeIs[type[SupportsMambaPrefixCaching]]: ...


def supports_mamba_prefix_caching(
    model: type[object] | object,
) -> TypeIs[type[SupportsMambaPrefixCaching]] | TypeIs[SupportsMambaPrefixCaching]:
    return getattr(model, "supports_mamba_prefix_caching", False)


746
747
748
749
750
751
752
753
754
@runtime_checkable
class SupportsCrossEncoding(Protocol):
    """The interface required for all models that support cross encoding."""

    supports_cross_encoding: ClassVar[Literal[True]] = True


@overload
def supports_cross_encoding(
755
756
    model: type[object],
) -> TypeIs[type[SupportsCrossEncoding]]: ...
757
758
759


@overload
760
def supports_cross_encoding(model: object) -> TypeIs[SupportsCrossEncoding]: ...
761
762
763


def _supports_cross_encoding(
764
765
    model: type[object] | object,
) -> TypeIs[type[SupportsCrossEncoding]] | TypeIs[SupportsCrossEncoding]:
766
    return getattr(model, "supports_cross_encoding", False)
767
768
769


def supports_cross_encoding(
770
771
    model: type[object] | object,
) -> TypeIs[type[SupportsCrossEncoding]] | TypeIs[SupportsCrossEncoding]:
772
    return is_pooling_model(model) and _supports_cross_encoding(model)
773
774


775
776
777
class SupportsQuant:
    """The interface required for all models that support quantization."""

778
779
780
    hf_to_vllm_mapper: ClassVar[WeightsMapper | None] = None
    packed_modules_mapping: ClassVar[dict[str, list[str]] | None] = None
    quant_config: QuantizationConfig | None = None
781

782
    def __new__(cls, *args, **kwargs) -> Self:
783
        instance = super().__new__(cls)
784
785

        # find config passed in arguments
786
787
        quant_config = cls._find_quant_config(*args, **kwargs)
        if quant_config is not None:
788
            # attach config to model for general use
789
            instance.quant_config = quant_config
790
791

            # apply model mappings to config for proper config-model matching
792
793
794
            if (hf_to_vllm_mapper := instance.hf_to_vllm_mapper) is not None:
                instance.quant_config.apply_vllm_mapper(hf_to_vllm_mapper)
            if instance.packed_modules_mapping is not None:
795
                instance.quant_config.packed_modules_mapping.update(
796
797
                    instance.packed_modules_mapping
                )
798

799
800
801
        return instance

    @staticmethod
802
    def _find_quant_config(*args, **kwargs) -> QuantizationConfig | None:
803
        """Find quant config passed through model constructor args"""
804
805
806
807
808
809
810
811
812
813
814
815
816
        from vllm.config import VllmConfig  # avoid circular import

        args_values = list(args) + list(kwargs.values())
        for arg in args_values:
            if isinstance(arg, VllmConfig):
                return arg.quant_config

            if isinstance(arg, QuantizationConfig):
                return arg

        return None


817
818
819
@runtime_checkable
class SupportsTranscription(Protocol):
    """The interface required for all models that support transcription."""
820

821
822
    # Mapping from ISO639_1 language codes: language names
    supported_languages: ClassVar[Mapping[str, str]]
823
824
825

    supports_transcription: ClassVar[Literal[True]] = True

826
827
828
829
830
831
    supports_transcription_only: ClassVar[bool] = False
    """
    Transcription models can opt out of text generation by setting this to
    `True`.
    """

832
833
834
835
836
837
838
839
840
    def __init_subclass__(cls, **kwargs):
        super().__init_subclass__(**kwargs)
        # language codes in supported_languages
        # that don't exist in the full language map
        invalid = set(cls.supported_languages) - set(LANGUAGES.keys())
        if invalid:
            raise ValueError(
                f"{cls.__name__}.supported_languages contains invalid "
                f"language codes: {sorted(invalid)}\n. "
841
842
                f"Valid choices are: {sorted(LANGUAGES.keys())}"
            )
843

844
    @classmethod
845
846
847
848
849
    def get_generation_prompt(
        cls,
        audio: np.ndarray,
        stt_config: SpeechToTextConfig,
        model_config: ModelConfig,
850
        language: str | None,
851
852
        task_type: Literal["transcribe", "translate"],
        request_prompt: str,
853
        to_language: str | None,
854
    ) -> PromptType:
855
856
857
        """Get the prompt for the ASR model.
        The model has control over the construction, as long as it
        returns a valid PromptType."""
858
859
860
        ...

    @classmethod
861
862
    def get_other_languages(cls) -> Mapping[str, str]:
        # other possible language codes from the whisper map
863
        return {k: v for k, v in LANGUAGES.items() if k not in cls.supported_languages}
864
865

    @classmethod
866
    def validate_language(cls, language: str | None) -> str | None:
867
        """
868
869
870
        Ensure the language specified in the transcription request
        is a valid ISO 639-1 language code. If the request language is
        valid, but not natively supported by the model, trigger a
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
        warning (but not an exception).
        """
        if language is None or language in cls.supported_languages:
            return language
        elif language in cls.get_other_languages():
            logger.warning(
                "Language %r is not natively supported by %s; "
                "results may be less accurate. Supported languages: %r",
                language,
                cls.__name__,
                list(cls.supported_languages.keys()),
            )
            return language
        else:
            raise ValueError(
                f"Unsupported language: {language!r}.  Must be one of "
887
888
                f"{list(cls.supported_languages.keys())}."
            )
889

890
891
    @classmethod
    def get_speech_to_text_config(
892
893
        cls, model_config: ModelConfig, task_type: Literal["transcribe", "translate"]
    ) -> SpeechToTextConfig:
894
895
896
897
        """Get the speech to text config for the ASR model."""
        ...

    @classmethod
898
899
900
901
902
    def get_num_audio_tokens(
        cls,
        audio_duration_s: float,
        stt_config: SpeechToTextConfig,
        model_config: ModelConfig,
903
    ) -> int | None:
904
        """
905
        Map from audio duration to number of audio tokens produced by the ASR
906
907
908
909
910
        model, without running a forward pass.
        This is used for estimating the amount of processing for this audio.
        """
        return None

911
912
913

@overload
def supports_transcription(
914
915
    model: type[object],
) -> TypeIs[type[SupportsTranscription]]: ...
916
917
918


@overload
919
def supports_transcription(model: object) -> TypeIs[SupportsTranscription]: ...
920
921
922


def supports_transcription(
923
924
    model: type[object] | object,
) -> TypeIs[type[SupportsTranscription]] | TypeIs[SupportsTranscription]:
925
    return getattr(model, "supports_transcription", False)
926
927


928
929
@runtime_checkable
class SupportsEagle3(Protocol):
930
    """The interface required for models that support
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
    EAGLE3 speculative decoding."""

    supports_eagle3: ClassVar[Literal[True]] = True
    """
    A flag that indicates this model supports EAGLE3 
    speculative decoding.

    Note:
        There is no need to redefine this flag if this class is in the
        MRO of your model class.
    """

    def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
        """
        Set which layers should output auxiliary
        hidden states for EAGLE3.
947

948
949
        Args:
            layers: Tuple of layer indices that should output auxiliary
950
                hidden states.
951
952
953
954
955
956
957
        """
        ...

    def get_eagle3_aux_hidden_state_layers(self) -> tuple[int, ...]:
        """
        Get the layer indices that should output auxiliary hidden states
        for EAGLE3.
958

959
960
961
962
963
964
965
        Returns:
            Tuple of layer indices for auxiliary hidden state outputs.
        """
        ...


@overload
966
def supports_eagle3(model: type[object]) -> TypeIs[type[SupportsEagle3]]: ...
967
968
969


@overload
970
def supports_eagle3(model: object) -> TypeIs[SupportsEagle3]: ...
971
972
973


def supports_eagle3(
974
975
    model: type[object] | object,
) -> TypeIs[type[SupportsEagle3]] | TypeIs[SupportsEagle3]:
976
    return isinstance(model, SupportsEagle3)
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994


@runtime_checkable
class SupportsMRoPE(Protocol):
    """The interface required for all models that support M-RoPE."""

    supports_mrope: ClassVar[Literal[True]] = True
    """
    A flag that indicates this model supports M-RoPE.
    
    Note:
        There is no need to redefine this flag if this class is in the
        MRO of your model class.
    """

    def get_mrope_input_positions(
        self,
        input_tokens: list[int],
995
        mm_features: list["MultiModalFeatureSpec"],
996
997
998
    ) -> tuple[torch.Tensor, int]:
        """
        Get M-RoPE input positions and delta value for this specific model.
999

1000
1001
        This method should be implemented by each model that supports M-RoPE
        to provide model-specific logic for computing input positions.
1002

1003
1004
        Args:
            input_tokens: List of input token IDs
1005
            mm_features: Information about each multi-modal data item
1006

1007
        Returns:
1008
1009
            Tuple of `(llm_positions, mrope_position_delta)`
            - llm_positions: Tensor of shape `[3, num_tokens]` with T/H/W positions
1010
1011
1012
1013
1014
1015
            - mrope_position_delta: Delta for position calculations
        """
        ...


@overload
1016
def supports_mrope(model: type[object]) -> TypeIs[type[SupportsMRoPE]]: ...
1017
1018
1019


@overload
1020
def supports_mrope(model: object) -> TypeIs[SupportsMRoPE]: ...
1021
1022
1023


def supports_mrope(
1024
1025
    model: type[object] | object,
) -> TypeIs[type[SupportsMRoPE]] | TypeIs[SupportsMRoPE]:
1026
    return isinstance(model, SupportsMRoPE)