Skip to content

vllm.model_executor.models.internvl

BaseInternVLDummyInputsBuilder

Bases: BaseDummyInputsBuilder[_I]

Basic image-only DummyInputsBuilder for InternVL-style models.

Source code in vllm/model_executor/models/internvl.py
class BaseInternVLDummyInputsBuilder(BaseDummyInputsBuilder[_I]):
    """Basic image-only DummyInputsBuilder for InternVL-style models."""

    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        num_images = mm_counts.get("image", 0)

        return "<image>" * num_images

    def get_dummy_mm_data(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
        mm_options: Mapping[str, BaseDummyOptions],
    ) -> MultiModalDataDict:
        target_width, target_height = self.info.get_image_size_with_most_features()
        num_images = mm_counts.get("image", 0)

        image_overrides = mm_options.get("image")

        return {
            "image": self._get_dummy_images(
                width=target_width,
                height=target_height,
                num_images=num_images,
                overrides=image_overrides,
            )
        }

BaseInternVLMultiModalProcessor

Bases: BaseMultiModalProcessor[_I]

Basic image-only MultiModalProcessor for InternVL-style models.

Source code in vllm/model_executor/models/internvl.py
class BaseInternVLMultiModalProcessor(BaseMultiModalProcessor[_I]):
    """Basic image-only MultiModalProcessor for InternVL-style models."""

    def _call_hf_processor(
        self,
        prompt: str,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
        tok_kwargs: Mapping[str, object],
    ) -> BatchFeature:
        processed_outputs = super()._call_hf_processor(
            prompt=prompt,
            mm_data=mm_data,
            mm_kwargs=mm_kwargs,
            tok_kwargs=tok_kwargs,
        )

        hf_processor = self.info.get_hf_processor(**mm_kwargs)
        image_token_id = hf_processor.ctx_image_token_id

        # Since there may be extra tokens in the feature placeholders,
        # we need to pass the image token ID to the model to select the
        # tokens to merge from the vision encoder outputs
        processed_outputs["image_token_id"] = torch.tensor(image_token_id)

        return processed_outputs

    def _get_image_fields_config(self, hf_inputs: BatchFeature):
        image_num_patches = hf_inputs.get("image_num_patches", torch.empty(0))
        num_images = len(image_num_patches)

        return dict(
            pixel_values_flat=MultiModalFieldConfig.flat_from_sizes(
                "image", image_num_patches
            ),
            image_num_patches=MultiModalFieldConfig.batched("image"),
            image_embeds=MultiModalFieldConfig.batched("image"),
            image_token_id=MultiModalFieldConfig.shared("image", num_images),
        )

    def _get_mm_fields_config(
        self,
        hf_inputs: BatchFeature,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
        return self._get_image_fields_config(hf_inputs)

    def _get_prompt_repl_image(
        self,
        mm_items: MultiModalDataItems,
        hf_processor: InternVLProcessor,
        out_mm_data: BatchedTensorInputs,
    ):
        if "image_num_patches" in out_mm_data:
            image_num_patches = out_mm_data["image_num_patches"]
            assert isinstance(image_num_patches, torch.Tensor)
            image_num_patches = image_num_patches.tolist()
        elif "image_embeds" in out_mm_data:
            # TODO: Use image size information in dictionary embedding inputs
            # to compute num_patches (similar to Qwen2-VL)
            image_num_patches = [None] * len(out_mm_data["image_embeds"])
        else:
            image_num_patches = []

        def get_replacement_internvl(item_idx: int):
            images = mm_items.get_items(
                "image", (ImageEmbeddingItems, ImageProcessorItems)
            )

            if isinstance(images, ImageEmbeddingItems):
                feature_size = images.get_feature_size(item_idx)
            else:
                image_size = images.get_image_size(item_idx)
                feature_size = self.info.get_num_image_tokens(
                    image_width=image_size.width,
                    image_height=image_size.height,
                    processor=hf_processor,
                )

            num_patches = image_num_patches[item_idx]
            if num_patches is not None:
                assert isinstance(num_patches, int)

            return hf_processor.get_image_repl(num_patches, num_features=feature_size)

        return PromptReplacement(
            modality="image",
            target="<image>",
            replacement=get_replacement_internvl,
        )

    def _get_prompt_updates(
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
        out_mm_kwargs: MultiModalKwargsItems,
    ) -> Sequence[PromptUpdate]:
        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
        out_mm_data = out_mm_kwargs.get_data()

        return [
            self._get_prompt_repl_image(mm_items, hf_processor, out_mm_data),
        ]

BaseInternVLProcessingInfo

Bases: BaseProcessingInfo

Basic image-only ProcessingInfo for InternVL-style models.

Source code in vllm/model_executor/models/internvl.py
class BaseInternVLProcessingInfo(BaseProcessingInfo):
    """Basic image-only ProcessingInfo for InternVL-style models."""

    @abstractmethod
    def get_hf_processor(self, **kwargs: object) -> InternVLProcessor:
        raise NotImplementedError

    def get_supported_mm_limits(self) -> Mapping[str, int | None]:
        return {"image": None}

    def get_num_image_tokens(
        self,
        *,
        image_width: int,
        image_height: int,
        processor: InternVLProcessor,
    ) -> int:
        return processor.get_num_image_tokens(
            image_width=image_width,
            image_height=image_height,
        )

    def get_image_size_with_most_features(self) -> ImageSize:
        processor = self.get_hf_processor()
        image_processor = processor.image_processor

        base_size = image_processor.image_size
        target_ratios = processor.resolve_target_ratios()

        largest_feature_size, largest_feature_pinpoint = 0, None
        for wr, hr in target_ratios:
            width, height = base_size * wr, base_size * hr

            feat_size = self.get_num_image_tokens(
                image_width=width,
                image_height=height,
                processor=processor,
            )
            if feat_size > largest_feature_size:
                largest_feature_size = feat_size
                largest_feature_pinpoint = ImageSize(width=width, height=height)

        if largest_feature_size == 0 or largest_feature_pinpoint is None:
            raise ValueError("Cannot have a largest feature size of 0!")

        return largest_feature_pinpoint

    def get_max_image_tokens(self) -> int:
        processor = self.get_hf_processor()
        target_width, target_height = self.get_image_size_with_most_features()

        return self.get_num_image_tokens(
            image_width=target_width,
            image_height=target_height,
            processor=processor,
        )

InternVLChatModel

Bases: Module, SupportsMultiModal, SupportsPP, SupportsLoRA, SupportsEncoderCudaGraph

Source code in vllm/model_executor/models/internvl.py
 540
 541
 542
 543
 544
 545
 546
 547
 548
 549
 550
 551
 552
 553
 554
 555
 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
 586
 587
 588
 589
 590
 591
 592
 593
 594
 595
 596
 597
 598
 599
 600
 601
 602
 603
 604
 605
 606
 607
 608
 609
 610
 611
 612
 613
 614
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 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
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 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
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
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
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
@MULTIMODAL_REGISTRY.register_processor(
    InternVLMultiModalProcessor,
    info=InternVLProcessingInfo,
    dummy_inputs=InternVLDummyInputsBuilder,
)
class InternVLChatModel(
    nn.Module,
    SupportsMultiModal,
    SupportsPP,
    SupportsLoRA,
    SupportsEncoderCudaGraph,
):
    supports_encoder_tp_data = True
    supports_encoder_cudagraph: ClassVar[Literal[True]] = True

    @classmethod
    def get_placeholder_str(cls, modality: str, i: int) -> str | None:
        if modality.startswith("image"):
            return "<image>"
        if modality.startswith("video"):
            return "<video>"

        raise ValueError("Only image or video modality is supported")

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
        super().__init__()

        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config

        self.config = config
        self.multimodal_config = multimodal_config
        self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
        self._patch_quant_config(config, quant_config)

        image_size = config.force_image_size or config.vision_config.image_size
        patch_size = config.vision_config.patch_size
        self.patch_size = patch_size
        self.patch_tokens = (image_size // patch_size) ** 2
        self.num_image_token = int(self.patch_tokens * (config.downsample_ratio**2))
        self.downsample_ratio = config.downsample_ratio
        self.ps_version = config.ps_version

        llm_arch_name = config.text_config.architectures[0]
        self.is_mono = llm_arch_name == "InternLM2VEForCausalLM"

        with self._mark_tower_model(vllm_config, {"image", "video"}):
            self.vision_model = self._init_vision_model(
                config,
                quant_config=quant_config,
                is_mono=self.is_mono,
                prefix=maybe_prefix(prefix, "vision_model"),
            )
            self.mlp1 = self._init_mlp1(config)

        with self._mark_language_model(vllm_config):
            self.language_model = init_vllm_registered_model(
                vllm_config=vllm_config,
                hf_config=config.text_config,
                prefix=maybe_prefix(prefix, "language_model"),
            )

        self.img_context_token_id = None
        self.video_context_token_id = None

        self.visual_token_mask = None
        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors
        )

    def _patch_quant_config(
        self, config: PretrainedConfig, quant_config: QuantizationConfig
    ):
        # the awq models from OpenGVLab missing `modules_to_not_convert`
        # patch the quant_config to add `modules_to_not_convert` back
        if isinstance(quant_config, AWQConfig):
            text_config = config.text_config
            llm_quant_config = getattr(text_config, "quantization_config", None)
            if (not quant_config.modules_to_not_convert) and (
                llm_quant_config is not None
            ):
                quant_config.modules_to_not_convert.append("vision_model")

    def _init_vision_model(
        self,
        config: PretrainedConfig,
        quant_config: QuantizationConfig | None,
        *,
        is_mono: bool,
        prefix: str,
    ):
        if not is_mono:
            vision_feature_layer = config.select_layer
            if vision_feature_layer < 0:
                num_hidden_layers = (
                    config.vision_config.num_hidden_layers + vision_feature_layer + 1
                )
            else:
                num_hidden_layers = vision_feature_layer + 1

            return InternVisionModel(
                config.vision_config,
                quant_config=quant_config,
                num_hidden_layers_override=num_hidden_layers,
                prefix=prefix,
            )
        else:
            return InternVisionPatchModel(config.vision_config)

    def _init_mlp1(self, config: PretrainedConfig) -> nn.Module:
        vit_hidden_size = config.vision_config.hidden_size
        llm_hidden_size = config.text_config.hidden_size

        return nn.Sequential(
            nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2),
            nn.Linear(
                vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size
            ),
            nn.GELU(),
            nn.Linear(llm_hidden_size, llm_hidden_size),
        )

    def pixel_shuffle(self, x, scale_factor=0.5):
        n, w, h, c = x.size()
        # N, W, H, C --> N, W, H * scale, C // scale
        x = x.view(n, w, int(h * scale_factor), int(c / scale_factor))
        # N, W, H * scale, C // scale --> N, H * scale, W, C // scale
        x = x.permute(0, 2, 1, 3).contiguous()
        x = x.view(
            n,
            int(h * scale_factor),
            int(w * scale_factor),
            int(c / (scale_factor * scale_factor)),
        )
        if self.ps_version == "v1":
            pass
        else:
            x = x.permute(0, 2, 1, 3).contiguous()
        return x

    def extract_feature(self, pixel_values: torch.Tensor) -> torch.Tensor:
        vit_embeds = self.vision_model(pixel_values=pixel_values)
        vit_embeds = vit_embeds[:, 1:, :]

        h = w = int(vit_embeds.shape[1] ** 0.5)
        vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
        vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio)
        vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1])
        vit_embeds = self.mlp1(vit_embeds)
        return vit_embeds

    def _parse_and_validate_image_input(
        self, **kwargs: object
    ) -> InternVLImageInputs | None:
        pixel_values_flat = kwargs.pop("pixel_values_flat", None)
        image_num_patches = kwargs.pop("image_num_patches", None)
        image_embeds = kwargs.pop("image_embeds", None)

        if pixel_values_flat is None and image_embeds is None:
            return None

        if image_embeds is not None:
            return InternVLImageEmbeddingInputs(
                type="image_embeds",
                data=image_embeds,
            )

        image_token_id = kwargs["image_token_id"]
        if isinstance(image_token_id, torch.Tensor):
            image_token_id = image_token_id.flatten().unique().item()

        assert isinstance(image_token_id, int)
        self.img_context_token_id = image_token_id

        if pixel_values_flat is not None:
            expected_h = expected_w = self.config.vision_config.image_size
            resolve_bindings = {"h": expected_h, "w": expected_w}

            return InternVLImagePixelInputs(
                type="pixel_values",
                pixel_values_flat=pixel_values_flat,
                num_patches=image_num_patches,
                resolve_bindings=resolve_bindings,
            )

        raise AssertionError("This line should be unreachable.")

    def _parse_and_validate_video_input(
        self, **kwargs: object
    ) -> InternVLVideoPixelInputs | None:
        pixel_values_flat_video = kwargs.pop("pixel_values_flat_video", None)
        video_num_patches = kwargs.pop("video_num_patches", None)
        video_embeds = kwargs.pop("image_embeds", None)

        if pixel_values_flat_video is None and video_embeds is None:
            return None

        if video_embeds is not None:
            return InternVLVideoEmbeddingInputs(
                type="video_embeds",
                data=video_embeds,
            )

        video_token_id = kwargs["video_token_id"]
        if isinstance(video_token_id, torch.Tensor):
            video_token_id = video_token_id.flatten().unique().item()

        assert isinstance(video_token_id, int)
        self.video_context_token_id = video_token_id

        if pixel_values_flat_video is not None:
            expected_h = expected_w = self.config.vision_config.image_size
            resolve_bindings = {"h": expected_h, "w": expected_w}

            return InternVLVideoPixelInputs(
                type="pixel_values_videos",
                pixel_values_flat=pixel_values_flat_video,
                num_patches=video_num_patches,
                resolve_bindings=resolve_bindings,
            )

        raise AssertionError("This line should be unreachable.")

    def _process_vision_input(
        self,
        image_input: InternVLImageInputs | InternVLVideoInputs,
    ) -> tuple[torch.Tensor, ...]:
        if (
            image_input["type"] == "image_embeds"
            or image_input["type"] == "video_embeds"
        ):
            return image_input["data"]

        image_embeds = self.extract_feature(image_input["pixel_values_flat"])

        num_patches = image_input["num_patches"]

        # Only one image in the current batch
        if len(num_patches) == 1:
            return (image_embeds.view(-1, self.config.text_config.hidden_size),)

        # NOTE: Image embeddings are split into separate tensors for each image
        # by the size of each embedding.
        feature_size = image_embeds.shape[1]
        image_embeds = image_embeds.view(-1, self.config.text_config.hidden_size)
        image_feature_sizes = [
            num_patches * feature_size for num_patches in num_patches
        ]
        return image_embeds.split(image_feature_sizes)

    def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
        modalities = {}

        # Preserve the order of modalities if there are multiple of them
        # from the order of kwargs.
        for input_key in kwargs:
            if (
                input_key in ("pixel_values_flat", "image_embeds")
                and "images" not in modalities
            ):
                modalities["images"] = self._parse_and_validate_image_input(**kwargs)
            if input_key in ("pixel_values_flat_video",) and "videos" not in modalities:
                modalities["videos"] = self._parse_and_validate_video_input(**kwargs)

        return modalities

    def _set_visual_token_mask(self, input_ids: torch.Tensor) -> None:
        if self.is_mono:
            assert self.img_context_token_id is not None
            self.visual_token_mask = (input_ids == self.img_context_token_id).reshape(
                -1, 1
            )
        else:
            self.visual_token_mask = None

    def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
        modalities = self._parse_and_validate_multimodal_inputs(**kwargs)
        if not modalities:
            return []

        # The result multimodal_embeddings is tuple of tensors, with each
        # tensor correspoending to a multimodal data item (image or video).
        multimodal_embeddings: tuple[torch.Tensor, ...] = ()

        # NOTE: It is important to iterate over the keys in this dictionary
        # to preserve the order of the modalities.
        for modality in modalities:
            if modality == "images":
                image_input = modalities["images"]
                image_embeddings = self._process_vision_input(image_input)
                multimodal_embeddings += tuple(image_embeddings)
            if modality == "videos":
                video_input = modalities["videos"]
                video_embeddings = self._process_vision_input(video_input)
                multimodal_embeddings += tuple(video_embeddings)

        return multimodal_embeddings

    def embed_input_ids(
        self,
        input_ids: torch.Tensor,
        multimodal_embeddings: MultiModalEmbeddings | None = None,
        *,
        is_multimodal: torch.Tensor | None = None,
    ) -> torch.Tensor:
        if multimodal_embeddings is not None and len(multimodal_embeddings) > 0:
            self._set_visual_token_mask(input_ids)

        # This is to satisfy the type checker for each overload
        if multimodal_embeddings is None or is_multimodal is None:
            return super().embed_input_ids(input_ids)

        return super().embed_input_ids(
            input_ids,
            multimodal_embeddings=multimodal_embeddings,
            is_multimodal=is_multimodal,
        )

    def forward(
        self,
        input_ids: torch.Tensor | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
        **kwargs: object,
    ) -> IntermediateTensors:
        if intermediate_tensors is not None:
            inputs_embeds = None

        forward_kwargs = {
            "input_ids": input_ids,
            "positions": positions,
            "intermediate_tensors": intermediate_tensors,
            "inputs_embeds": inputs_embeds,
        }

        # Only required if the model is mono-architecture
        if self.visual_token_mask is not None:
            forward_kwargs.update({"visual_token_mask": self.visual_token_mask})
            self.visual_token_mask = None

        hidden_states = self.language_model.model(**forward_kwargs)
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor | None:
        return self.language_model.compute_logits(hidden_states)

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        # unused modules appear in OpenGVLab/InternVideo2_5_Chat_8B
        skip_prefixes = [
            "action_embed",
            "temporal_embed",
            "track_embed",
            "track_embed_decoder",
            "box_token",
            "cg_criterion",
            "cg_model",
            "loc_encoder",
            "loc_decoder",
            "sam",
            "temporal_token",
            "track_token",
        ]
        loader = AutoWeightsLoader(self, skip_prefixes=skip_prefixes)
        return loader.load_weights(weights)

    def get_mm_mapping(self) -> MultiModelKeys:
        """
        Get the module prefix in multimodal models
        """
        return MultiModelKeys.from_string_field(
            language_model="language_model",
            connector="mlp1",
            tower_model="vision_model",
        )

    def get_num_mm_encoder_tokens(self, num_image_tokens: int) -> int:
        if num_image_tokens <= 0 or self.num_image_token <= 0:
            return 0

        num_patches = num_image_tokens // self.num_image_token
        return num_patches * (self.patch_tokens + 1)

    def get_num_mm_connector_tokens(self, num_vision_tokens: int) -> int:
        if num_vision_tokens <= 0 or self.num_image_token <= 0:
            return 0

        num_patches = num_vision_tokens // (self.patch_tokens + 1)
        return num_patches * self.num_image_token

    # -- SupportsEncoderCudaGraph protocol methods --

    def get_encoder_cudagraph_config(self):
        from vllm.v1.worker.encoder_cudagraph_defs import EncoderCudaGraphConfig

        return EncoderCudaGraphConfig(
            modalities=["image", "video"],
            input_key_by_modality={
                "image": "pixel_values_flat",
                "video": "pixel_values_flat_video",
            },
            # InternVision uses standard ViT attention (no rotary embeddings,
            # no variable-length sequence metadata), so no extra buffers needed.
            buffer_keys=[],
            out_hidden_size=self.config.text_config.hidden_size,
        )

    def get_input_modality(
        self,
        mm_kwargs: dict[str, Any],
    ) -> str:
        if "pixel_values_flat" in mm_kwargs:
            return "image"
        return "video"

    def get_max_frames_per_video(self) -> int:
        # InternVL has no attention-metadata buffers that depend on frame
        # count (buffer_keys=[]), so any value is safe. Return 1.
        return 1

    def get_encoder_cudagraph_budget_range(
        self,
        vllm_config: "VllmConfig",
    ) -> tuple[int, int]:
        # Min: 1 tile → num_image_token output tokens.
        min_budget = self.num_image_token
        max_budget = min(
            vllm_config.scheduler_config.max_num_batched_tokens,
            vllm_config.model_config.max_model_len,
        )
        return (min_budget, max_budget)

    def _get_internvl_patches_list(
        self,
        mm_kwargs: dict[str, Any],
    ) -> list[int]:
        """Return per-item tile counts as a plain list of ints."""
        if self.get_input_modality(mm_kwargs) == "image":
            patches = mm_kwargs.get("image_num_patches", [])
        else:
            patches = mm_kwargs.get("video_num_patches", [])
        if isinstance(patches, torch.Tensor):
            return patches.tolist()
        return [int(n) for n in patches]

    def get_encoder_cudagraph_num_items(
        self,
        mm_kwargs: dict[str, Any],
    ) -> int:
        return len(self._get_internvl_patches_list(mm_kwargs))

    def get_encoder_cudagraph_per_item_output_tokens(
        self,
        mm_kwargs: dict[str, Any],
    ) -> list[int]:
        return [
            n * self.num_image_token for n in self._get_internvl_patches_list(mm_kwargs)
        ]

    def get_encoder_cudagraph_per_item_input_sizes(
        self,
        mm_kwargs: dict[str, Any],
    ) -> list[int]:
        return self._get_internvl_patches_list(mm_kwargs)

    def select_encoder_cudagraph_items(
        self,
        mm_kwargs: dict[str, Any],
        indices: list[int],
    ) -> dict[str, Any]:
        modality = self.get_input_modality(mm_kwargs)
        pv_key = (
            "pixel_values_flat" if modality == "image" else "pixel_values_flat_video"
        )
        patches_key = (
            "image_num_patches" if modality == "image" else "video_num_patches"
        )

        pixel_values = mm_kwargs[pv_key]
        patches_list = self._get_internvl_patches_list(mm_kwargs)

        if len(indices) == 0:
            return {pv_key: pixel_values[:0], patches_key: []}

        # Compute cumulative tile offsets for slicing pixel_values.
        cum_patches = [0]
        for n in patches_list:
            cum_patches.append(cum_patches[-1] + n)

        selected_pv = torch.cat(
            [pixel_values[cum_patches[i] : cum_patches[i + 1]] for i in indices]
        )
        selected_patches = [patches_list[i] for i in indices]

        return {pv_key: selected_pv, patches_key: selected_patches}

    def prepare_encoder_cudagraph_capture_inputs(
        self,
        token_budget: int,
        max_batch_size: int,
        max_frames_per_batch: int,
        device: torch.device,
        dtype: torch.dtype,
    ):
        from vllm.v1.worker.encoder_cudagraph_defs import (
            EncoderCudaGraphCaptureInputs,
        )

        # Size the buffer to hold the maximum possible tiles for this budget.
        total_tiles = max(token_budget // self.num_image_token, 1)
        image_size = self.config.vision_config.image_size

        dummy_pixel_values = torch.randn(
            total_tiles, 3, image_size, image_size, device=device, dtype=dtype
        )
        mm_kwargs = {
            "pixel_values_flat": dummy_pixel_values,
            # Single dummy item consuming all tiles; not used inside
            # extract_feature, only needed for structural consistency.
            "image_num_patches": [total_tiles],
        }

        return EncoderCudaGraphCaptureInputs(mm_kwargs=mm_kwargs, buffers={})

    def prepare_encoder_cudagraph_replay_buffers(
        self,
        mm_kwargs: dict[str, Any],
        max_batch_size: int,
        max_frames_per_batch: int,
    ):
        from vllm.v1.worker.encoder_cudagraph_defs import (
            EncoderCudaGraphReplayBuffers,
        )

        # No metadata buffers required for InternVision.
        return EncoderCudaGraphReplayBuffers(buffers={})

    def encoder_cudagraph_forward(
        self,
        mm_kwargs: dict[str, Any],
        buffers: dict[str, torch.Tensor],
    ) -> torch.Tensor:
        # The graph is always captured with pixel_values_flat as the input
        # buffer. During video replay the manager copies video tiles into
        # this same buffer before calling graph.replay(), so we always read
        # from pixel_values_flat here.
        pixel_values = mm_kwargs["pixel_values_flat"]
        out = self.extract_feature(pixel_values)  # [N, num_image_token, H]
        return out.view(-1, self.config.text_config.hidden_size)

    def encoder_eager_forward(
        self,
        mm_kwargs: dict[str, Any],
    ) -> torch.Tensor:
        if self.get_input_modality(mm_kwargs) == "image":
            pixel_values = mm_kwargs["pixel_values_flat"]
        else:
            pixel_values = mm_kwargs["pixel_values_flat_video"]
        out = self.extract_feature(pixel_values)  # [N, num_image_token, H]
        return out.view(-1, self.config.text_config.hidden_size)

_get_internvl_patches_list

_get_internvl_patches_list(
    mm_kwargs: dict[str, Any],
) -> list[int]

Return per-item tile counts as a plain list of ints.

Source code in vllm/model_executor/models/internvl.py
def _get_internvl_patches_list(
    self,
    mm_kwargs: dict[str, Any],
) -> list[int]:
    """Return per-item tile counts as a plain list of ints."""
    if self.get_input_modality(mm_kwargs) == "image":
        patches = mm_kwargs.get("image_num_patches", [])
    else:
        patches = mm_kwargs.get("video_num_patches", [])
    if isinstance(patches, torch.Tensor):
        return patches.tolist()
    return [int(n) for n in patches]

get_mm_mapping

get_mm_mapping() -> MultiModelKeys

Get the module prefix in multimodal models

Source code in vllm/model_executor/models/internvl.py
def get_mm_mapping(self) -> MultiModelKeys:
    """
    Get the module prefix in multimodal models
    """
    return MultiModelKeys.from_string_field(
        language_model="language_model",
        connector="mlp1",
        tower_model="vision_model",
    )

InternVLDummyInputsBuilder

Bases: BaseInternVLDummyInputsBuilder[InternVLProcessingInfo]

InternVL DummyInputsBuilder extended for video support

Source code in vllm/model_executor/models/internvl.py
class InternVLDummyInputsBuilder(
    BaseInternVLDummyInputsBuilder[InternVLProcessingInfo]
):
    """InternVL DummyInputsBuilder extended for video support"""

    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        num_videos = mm_counts.get("video", 0)

        return super().get_dummy_text(mm_counts) + "<video>" * num_videos

    def get_dummy_mm_data(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
        mm_options: Mapping[str, BaseDummyOptions],
    ) -> MultiModalDataDict:
        dummy_image = super().get_dummy_mm_data(seq_len, mm_counts, mm_options)
        if self.info.ctx_video_token:
            config = self.info.get_hf_config()
            image_size: int = config.vision_config.image_size
            target_num_frames = self.info.get_num_frames_with_most_features(
                seq_len, mm_counts
            )
            num_videos = mm_counts.get("video", 0)
            video_overrides = mm_options.get("video")
            dummy_video = {
                "video": self._get_dummy_videos(
                    width=image_size,
                    height=image_size,
                    num_frames=target_num_frames,
                    num_videos=num_videos,
                    overrides=video_overrides,
                )
            }
        else:
            dummy_video = {}
        return {**dummy_image, **dummy_video}

InternVLImageEmbeddingInputs

Bases: TensorSchema

Dimensions
  • n: Number of images
  • f: Total image feature size
  • h: Hidden size (must match the hidden size of language model backbone)
Source code in vllm/model_executor/models/internvl.py
class InternVLImageEmbeddingInputs(TensorSchema):
    """
    Dimensions:
        - n: Number of images
        - f: Total image feature size
        - h: Hidden size (must match the hidden size of language model backbone)
    """

    type: Literal["image_embeds"]
    data: Annotated[torch.Tensor | list[torch.Tensor], TensorShape("n", "f", "h")]

InternVLImagePixelInputs

Bases: TensorSchema

Dimensions
  • bn: Batch size * number of images
  • bnp: Batch size * number of images * (1 + num_patches)
  • c: Number of channels (3)
  • h: Height of each image patch
  • w: Width of each image patch
Source code in vllm/model_executor/models/internvl.py
class InternVLImagePixelInputs(TensorSchema):
    """
    Dimensions:
        - bn: Batch size * number of images
        - bnp: Batch size * number of images * (1 + num_patches)
        - c: Number of channels (3)
        - h: Height of each image patch
        - w: Width of each image patch
    """

    type: Literal["pixel_values"]
    pixel_values_flat: Annotated[torch.Tensor, TensorShape("bnp", 3, "h", "w")]
    num_patches: Annotated[torch.Tensor, TensorShape("bn")]

InternVLMultiModalProcessor

Bases: BaseInternVLMultiModalProcessor[InternVLProcessingInfo]

InternVL MultiModalProcessor extended for video support

Source code in vllm/model_executor/models/internvl.py
class InternVLMultiModalProcessor(
    BaseInternVLMultiModalProcessor[InternVLProcessingInfo]
):
    """InternVL MultiModalProcessor extended for video support"""

    def _call_hf_processor(
        self,
        prompt: str,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
        tok_kwargs: Mapping[str, object],
    ) -> BatchFeature:
        processed_outputs = super()._call_hf_processor(
            prompt, mm_data, mm_kwargs, tok_kwargs
        )

        hf_processor = self.info.get_hf_processor(**mm_kwargs)
        if (video_token_id := hf_processor.ctx_video_token_id) is not None:
            processed_outputs["video_token_id"] = torch.tensor(video_token_id)

        return processed_outputs

    def _get_video_fields_config(self, hf_inputs: BatchFeature):
        video_num_patches = hf_inputs.get("video_num_patches", torch.empty(0))
        num_videos = len(video_num_patches)

        return dict(
            pixel_values_flat_video=MultiModalFieldConfig.flat_from_sizes(
                "video", video_num_patches
            ),
            video_num_patches=MultiModalFieldConfig.batched("video"),
            video_token_id=MultiModalFieldConfig.shared("video", num_videos),
        )

    def _get_mm_fields_config(
        self,
        hf_inputs: BatchFeature,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
        fields = self._get_image_fields_config(hf_inputs)
        if self.info.ctx_video_token:
            fields |= self._get_video_fields_config(hf_inputs)

        return fields

    def _get_prompt_repl_video(
        self,
        mm_items: MultiModalDataItems,
        hf_processor: InternVLProcessor,
        out_mm_data: BatchedTensorInputs,
    ):
        if "video_num_patches" in out_mm_data:
            video_num_patches = out_mm_data["video_num_patches"]
            assert isinstance(video_num_patches, torch.Tensor)
            video_num_patches = video_num_patches.tolist()
        else:
            video_num_patches = []

        def get_video_replacement_internvl(item_idx: int):
            num_patches = video_num_patches[item_idx]
            if num_patches is not None:
                assert isinstance(num_patches, int)

            return hf_processor.get_video_repl(num_patches)

        return PromptReplacement(
            modality="video",
            target="<video>",
            replacement=get_video_replacement_internvl,
        )

    def _get_prompt_updates(
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
        out_mm_kwargs: MultiModalKwargsItems,
    ) -> Sequence[PromptUpdate]:
        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
        out_mm_data = out_mm_kwargs.get_data()

        prompt_repls = [
            self._get_prompt_repl_image(mm_items, hf_processor, out_mm_data),
        ]
        if self.info.ctx_video_token is not None:
            prompt_repls.append(
                self._get_prompt_repl_video(mm_items, hf_processor, out_mm_data)
            )

        return prompt_repls

InternVLProcessingInfo

Bases: BaseInternVLProcessingInfo

InternVL ProcessingInfo extended for video processing

Source code in vllm/model_executor/models/internvl.py
class InternVLProcessingInfo(BaseInternVLProcessingInfo):
    """InternVL ProcessingInfo extended for video processing"""

    def get_image_processor(self, **kwargs):
        config = self.get_hf_config()
        vision_config = config.vision_config

        kwargs = self.ctx.get_merged_mm_kwargs(kwargs)
        kwargs.setdefault("image_size", vision_config.image_size)
        kwargs.setdefault("min_dynamic_patch", config.min_dynamic_patch)
        kwargs.setdefault("max_dynamic_patch", config.max_dynamic_patch)
        kwargs.setdefault("dynamic_image_size", config.dynamic_image_size)
        kwargs.setdefault("use_thumbnail", config.use_thumbnail)

        return InternVLImageProcessor(**kwargs)

    def get_video_processor(self, **kwargs):
        config = self.get_hf_config()
        vision_config = config.vision_config

        kwargs = self.ctx.get_merged_mm_kwargs(kwargs)
        kwargs.setdefault("image_size", vision_config.image_size)

        return InternVLVideoProcessor(**kwargs)

    @cached_property
    def ctx_video_token(self):
        text_model_type = self.get_hf_config().get_text_config().model_type
        ctx_video_token_map = {
            "qwen2": "<|video_pad|>",
            "qwen3": "<|video_pad|>",
            "qwen3_moe": "<|video_pad|>",
            "gpt_oss": "<|reserved_200000|>",
        }

        if text_model_type not in ctx_video_token_map:
            return None

        ctx_video_token = ctx_video_token_map[text_model_type]
        if ctx_video_token not in self.get_tokenizer().get_vocab():
            return None

        return ctx_video_token

    def get_hf_processor(self, **kwargs: object) -> InternVLProcessor:
        config = self.get_hf_config()
        vision_config = config.vision_config

        image_processor = self.get_image_processor(**kwargs)
        image_size = image_processor.image_size
        patch_size = vision_config.patch_size
        downsample_ratio = config.downsample_ratio
        image_seq_length = int((image_size // patch_size) ** 2 * (downsample_ratio**2))

        ctx_video_token = self.ctx_video_token
        video_processor = (
            self.get_video_processor(**kwargs) if ctx_video_token else None
        )

        return InternVLProcessor(
            tokenizer=self.get_tokenizer(),
            image_processor=image_processor,
            video_processor=video_processor,
            image_seq_length=image_seq_length,
            ctx_video_token=ctx_video_token,
        )

    def get_supported_mm_limits(self):
        video_limit = {"video": None} if self.ctx_video_token else {}
        return {**super().get_supported_mm_limits(), **video_limit}

    def get_num_frames_with_most_features(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> int:
        max_images = mm_counts.get("image", 0)
        max_videos = mm_counts.get("video", 0)

        processor = self.get_hf_processor()
        num_image_token = processor.image_seq_length

        max_image_tokens = self.get_max_image_tokens() * max_images
        max_total_frames = (seq_len - max_image_tokens) // num_image_token
        max_frames_per_video = max_total_frames // max(max_videos, 1)

        return max(max_frames_per_video, 1)

InternVLVideoEmbeddingInputs

Bases: TensorSchema

Dimensions
  • n: Number of videos
  • f: Total video feature size
  • h: Hidden size (must match the hidden size of language model backbone)
Source code in vllm/model_executor/models/internvl.py
class InternVLVideoEmbeddingInputs(TensorSchema):
    """
    Dimensions:
        - n: Number of videos
        - f: Total video feature size
        - h: Hidden size (must match the hidden size of language model backbone)
    """

    type: Literal["video_embeds"]
    data: Annotated[torch.Tensor | list[torch.Tensor], TensorShape("n", "f", "h")]

InternVLVideoPixelInputs

Bases: TensorSchema

Dimensions
  • bvf: Batch size * number of videos * num_frames
  • bn: Batch size * number of images
  • c: Number of channels (3)
  • h: Height of each video frame
  • w: Width of each video frame
Source code in vllm/model_executor/models/internvl.py
class InternVLVideoPixelInputs(TensorSchema):
    """
    Dimensions:
        - bvf: Batch size * number of videos * num_frames
        - bn: Batch size * number of images
        - c: Number of channels (3)
        - h: Height of each video frame
        - w: Width of each video frame
    """

    type: Literal["pixel_values_videos"]
    pixel_values_flat: Annotated[torch.Tensor, TensorShape("bvf", 3, "h", "w")]
    num_patches: Annotated[torch.Tensor, TensorShape("bn")]