vllm.model_executor.models.glm4_1v ¶

class Glm4vDummyInputsBuilder(BaseDummyInputsBuilder[Glm4vProcessingInfo]):

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

        hf_config = self.info.get_hf_config()
        hf_processor = self.info.get_hf_processor()
        tokenizer = self.info.get_tokenizer()

        image_token: str = hf_processor.image_token
        video_token_ids = [
            hf_config.video_start_token_id,
            hf_processor.video_token_id,
            hf_config.video_end_token_id,
        ]
        video_token = tokenizer.decode(video_token_ids)

        return image_token * num_images + video_token * num_videos

    def get_dummy_mm_data(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> MultiModalDataDict:
        num_images = mm_counts.get("image", 0)
        num_videos = mm_counts.get("video", 0)

        target_width, target_height = (
            self.info.get_image_size_with_most_features())
        target_num_frames = self.info.get_num_frames_with_most_features(
            seq_len, mm_counts)
        return {
            "image":
            self._get_dummy_images(width=target_width,
                                   height=target_height,
                                   num_images=num_images),
            "video":
            self._get_dummy_videos(
                width=target_width,
                height=target_height,
                num_frames=target_num_frames,
                num_videos=num_videos,
            ),
        }

    def _get_dummy_videos(
        self,
        *,
        width: int,
        height: int,
        num_frames: int,
        num_videos: int,
    ) -> list[VideoItem]:
        video = np.full((num_frames, width, height, 3), 255, dtype=np.uint8)
        video_items = []
        for i in range(num_videos):
            video_metadata = {
                "fps": 2.0,
                "duration": num_frames / 2.0,
                "total_num_frames": num_frames,
                "frames_indices": [i for i in range(num_frames)],
                "video_backend": "opencv",
                "do_sample_frames": False,
            }
            video_item = (video.copy(), video_metadata)
            video_items.append(video_item)

        return video_items

@MULTIMODAL_REGISTRY.register_processor(
    Glm4vMultiModalProcessor,
    info=Glm4vProcessingInfo,
    dummy_inputs=Glm4vDummyInputsBuilder,
)
class Glm4vForConditionalGeneration(nn.Module, SupportsMultiModal,
                                    SupportsLoRA, SupportsPP):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": ["gate_up_proj"]
    }

    # To ensure correct weight loading and mapping.
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={
            "lm_head.": "language_model.lm_head.",
            "model.language_model.": "language_model.model.",
            "model.visual.": "visual.",
        })

    supports_encoder_tp_data = True

    @classmethod
    def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
        if modality.startswith("image"):
            return "<|begin_of_image|><|image|><|end_of_image|>"
        if modality.startswith("video"):
            return "<|begin_of_video|><|video|><|end_of_video|>"

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

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        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.visual = Glm4vVisionTransformer(
            config.vision_config,
            norm_eps=getattr(config, "rms_norm_eps", 1e-5),
            quant_config=quant_config,
            prefix=maybe_prefix(prefix, "visual"),
            use_data_parallel=self.use_data_parallel,
        )

        if config.model_type == "glm4v":
            architectures = ["Glm4ForCausalLM"]
        elif config.model_type == "glm4v_moe":
            architectures = ["Glm4MoeForCausalLM"]
        else:
            architectures = None

        self.language_model = init_vllm_registered_model(
            vllm_config=vllm_config,
            hf_config=config.text_config,
            prefix=maybe_prefix(prefix, "language_model"),
            architectures=architectures)

        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors)

    def _validate_and_reshape_mm_tensor(self, mm_input: object,
                                        name: str) -> torch.Tensor:
        if not isinstance(mm_input, (torch.Tensor, list)):
            raise ValueError(
                f"Incorrect type of {name}. Got type: {type(mm_input)}")
        if isinstance(mm_input, torch.Tensor):
            if mm_input.ndim == 2:
                return mm_input
            if mm_input.ndim != 3:
                raise ValueError(f"{name} should be 2D or batched 3D tensor. "
                                 f"Got ndim: {mm_input.ndim} "
                                 f"(shape={mm_input.shape})")
            return mm_input.reshape(-1, mm_input.shape[-1])
        else:
            return torch.concat(mm_input)

    def _parse_and_validate_image_input(
            self, **kwargs: object) -> Optional[Glm4vImageInputs]:
        pixel_values = kwargs.pop("pixel_values", None)
        image_embeds = kwargs.pop("image_embeds", None)
        image_grid_thw = kwargs.pop("image_grid_thw", None)

        if pixel_values is None and image_embeds is None:
            return None

        if pixel_values is not None:
            pixel_values = self._validate_and_reshape_mm_tensor(
                pixel_values, "image pixel values")
            image_grid_thw = self._validate_and_reshape_mm_tensor(
                image_grid_thw, "image grid_thw")

            return Glm4vImagePixelInputs(
                type="pixel_values",
                pixel_values=pixel_values,
                image_grid_thw=image_grid_thw,
            )

        if image_embeds is not None:
            image_embeds = self._validate_and_reshape_mm_tensor(
                image_embeds, "image embeds")
            image_grid_thw = self._validate_and_reshape_mm_tensor(
                image_grid_thw, "image grid_thw")

            return Glm4vImageEmbeddingInputs(
                type="image_embeds",
                image_embeds=image_embeds,
                image_grid_thw=image_grid_thw,
            )

    def _parse_and_validate_video_input(
            self, **kwargs: object) -> Optional[Glm4vVideoInputs]:
        pixel_values_videos = kwargs.pop("pixel_values_videos", None)
        video_embeds = kwargs.pop("video_embeds", None)
        video_grid_thw = kwargs.pop("video_grid_thw", None)

        if pixel_values_videos is None and video_embeds is None:
            return None

        if pixel_values_videos is not None:
            pixel_values_videos = self._validate_and_reshape_mm_tensor(
                pixel_values_videos, "video pixel values")
            video_grid_thw = self._validate_and_reshape_mm_tensor(
                video_grid_thw, "video grid_thw")

            return Glm4vVideoPixelInputs(
                type="pixel_values_videos",
                pixel_values_videos=pixel_values_videos,
                video_grid_thw=video_grid_thw,
            )

        if video_embeds is not None:
            video_embeds = self._validate_and_reshape_mm_tensor(
                video_embeds, "video embeds")
            video_grid_thw = self._validate_and_reshape_mm_tensor(
                video_grid_thw, "video grid_thw")

            return Glm4vVideoEmbeddingInputs(
                type="video_embeds",
                video_embeds=video_embeds,
                video_grid_thw=video_grid_thw,
            )

    def _process_image_input(
            self, image_input: Glm4vImageInputs) -> tuple[torch.Tensor, ...]:
        grid_thw = image_input["image_grid_thw"]
        assert grid_thw.ndim == 2
        grid_thw_list = grid_thw.tolist()

        if image_input["type"] == "image_embeds":
            image_embeds = image_input["image_embeds"].type(self.visual.dtype)
        else:
            pixel_values = image_input["pixel_values"].type(self.visual.dtype)
            if self.use_data_parallel:
                return run_dp_sharded_mrope_vision_model(self.visual,
                                                         pixel_values,
                                                         grid_thw.tolist(),
                                                         rope_type="rope_3d")
            else:
                image_embeds = self.visual(pixel_values,
                                           grid_thw=grid_thw.tolist())
        merge_size = self.visual.spatial_merge_size
        sizes = (torch.tensor(grid_thw_list, dtype=torch.long).prod(-1) //
                 (merge_size * merge_size)).tolist()
        return image_embeds.split(sizes)

    def _process_video_input(
            self, video_input: Glm4vVideoInputs) -> tuple[torch.Tensor, ...]:
        grid_thw = video_input["video_grid_thw"]
        assert grid_thw.ndim == 2
        grid_thw_list = grid_thw.tolist()

        if video_input["type"] == "video_embeds":
            video_embeds = video_input["video_embeds"].type(self.visual.dtype)
        else:
            pixel_values_videos = video_input["pixel_values_videos"].type(
                self.visual.dtype)
            if self.use_data_parallel:
                return run_dp_sharded_mrope_vision_model(self.visual,
                                                         pixel_values_videos,
                                                         grid_thw.tolist(),
                                                         rope_type="rope_3d")
            else:
                video_embeds = self.visual(pixel_values_videos,
                                           grid_thw=grid_thw.tolist())
        # Split concatenated embeddings for each video item.
        merge_size = self.visual.spatial_merge_size
        sizes = (torch.tensor(grid_thw_list, dtype=torch.long).prod(-1) //
                 (merge_size * merge_size)).tolist()
        return video_embeds.split(sizes)

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

        # 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", "image_embeds")
                    and "image" not in mm_input_by_modality):
                mm_input_by_modality["image"] = (
                    self._parse_and_validate_image_input(**kwargs))
            if (input_key in ("pixel_values_videos", "video_embeds")
                    and "video" not in mm_input_by_modality):
                mm_input_by_modality["video"] = (
                    self._parse_and_validate_video_input(**kwargs))
        return mm_input_by_modality

    def get_language_model(self) -> torch.nn.Module:
        return self.language_model

    def get_multimodal_embeddings(
            self, **kwargs: object) -> Optional[MultiModalEmbeddings]:
        mm_input_by_modality = self._parse_and_validate_multimodal_inputs(
            **kwargs)
        if not mm_input_by_modality:
            return None

        # The result multimodal_embeddings is tuple of tensors, with each
        # tensor corresponding 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 mm_input_by_modality:
            multimodal_input = mm_input_by_modality[modality]
            if modality == "image":
                vision_embeddings = self._process_image_input(multimodal_input)
                multimodal_embeddings += vision_embeddings
            if modality == "video":
                video_embeddings = self._process_video_input(multimodal_input)
                multimodal_embeddings += video_embeddings
        return multimodal_embeddings

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        **kwargs: object,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        """Run forward pass for GLM-4V.

        Args:
            input_ids: Flattened (concatenated) input_ids corresponding to a
                batch.
            positions: Flattened (concatenated) position ids corresponding to a
                batch.
                **NOTE**: If mrope is enabled (default setting for GLM-4V
                opensource models), the shape will be `(3, seq_len)`,
                otherwise it will be `(seq_len,).
            intermediate_tensors: Optional intermediate tensors for pipeline
                parallelism.
            inputs_embeds: Optional pre-computed input embeddings.
            **kwargs: Additional keyword arguments.
        """
        if intermediate_tensors is not None:
            inputs_embeds = None

        hidden_states = self.language_model.model(
            input_ids=input_ids,
            positions=positions,
            intermediate_tensors=intermediate_tensors,
            inputs_embeds=inputs_embeds,
        )
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> Optional[torch.Tensor]:
        return self.language_model.compute_logits(hidden_states)

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

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

class Glm4vImageEmbeddingInputs(TensorSchema):
    """
    Dimensions:
        - f: Number of image features (varies based on image resolution)
        - h: Hidden size (must match language model backbone)
        - n: Number of images
        - g: Grid dimensions (3 for grid_t, grid_h, grid_w)
    """
    type: Literal["image_embeds"] = "image_embeds"

    image_embeds: Annotated[torch.Tensor, TensorShape("f", "h")]
    image_grid_thw: Annotated[torch.Tensor, TensorShape("n", 3)]

class Glm4vImagePixelInputs(TensorSchema):
    """
    Dimensions:
        - np: Number of patches
        - cpp: Number of channels * patch_size * patch_size
        - ni: Number of images
        - g: Grid dimensions (3 for grid_t, grid_h, grid_w)
    """
    type: Literal["pixel_values"] = "pixel_values"

    pixel_values: Annotated[torch.Tensor, TensorShape("np", "cpp")]
    image_grid_thw: Annotated[torch.Tensor, TensorShape("ni", 3)]