vllm.model_executor.models.qwen2_vl ¶

class Qwen2VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen2VLProcessingInfo]):

    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_processor = self.info.get_hf_processor()
        image_token: str = hf_processor.image_token
        video_token: str = hf_processor.video_token

        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,
            )
        }

@MULTIMODAL_REGISTRY.register_processor(Qwen2VLMultiModalProcessor,
                                        info=Qwen2VLProcessingInfo,
                                        dummy_inputs=Qwen2VLDummyInputsBuilder)
class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
                                      SupportsLoRA, SupportsPP, SupportsMRoPE):

    # To ensure correct weight loading and mapping.
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={
            # mapping for new names in checkpoint saved after transformers v4.52
            "model.language_model.": "language_model.model.",
            "model.visual.": "visual.",
            # mapping for original checkpoint
            "lm_head.": "language_model.lm_head.",
            "model.": "language_model.model.",
        })

    supports_encoder_tp_data = True

    def get_mrope_input_positions(
        self,
        input_tokens: list[int],
        hf_config: PretrainedConfig,
        image_grid_thw: Optional[Union[list[list[int]], torch.Tensor]],
        video_grid_thw: Optional[Union[list[list[int]], torch.Tensor]],
        second_per_grid_ts: Optional[list[float]] = None,
        context_len: int = 0,
        seq_len: Optional[int] = None,
        audio_feature_lengths: Optional[torch.Tensor] = None,
        use_audio_in_video: bool = False,
    ) -> tuple[torch.Tensor, int]:
        """Get M-RoPE input positions for Qwen2-VL model."""
        if image_grid_thw is None:
            image_grid_thw = []
        if video_grid_thw is None:
            video_grid_thw = []
        if second_per_grid_ts is None:
            second_per_grid_ts = []

        image_token_id = hf_config.image_token_id
        video_token_id = hf_config.video_token_id
        vision_start_token_id = hf_config.vision_start_token_id
        spatial_merge_size = hf_config.vision_config.spatial_merge_size
        tokens_per_second = getattr(hf_config.vision_config,
                                    "tokens_per_second", 1.0)

        input_tokens_tensor = torch.tensor(input_tokens)
        vision_start_indices = torch.argwhere(
            input_tokens_tensor == vision_start_token_id).squeeze(1)
        vision_tokens = input_tokens_tensor[vision_start_indices + 1]
        image_nums = (vision_tokens == image_token_id).sum()
        video_nums = (vision_tokens == video_token_id).sum()
        llm_pos_ids_list: list = []

        st = 0
        remain_images, remain_videos = image_nums, video_nums

        image_index, video_index = 0, 0
        for _ in range(image_nums + video_nums):
            video_second_per_grid_t = 0.0
            if remain_images > 0:
                try:
                    ed_image = input_tokens.index(image_token_id, st)
                except ValueError:
                    ed_image = len(input_tokens) + 1
            else:
                ed_image = len(input_tokens) + 1
            if remain_videos > 0:
                try:
                    ed_video = input_tokens.index(video_token_id, st)
                except ValueError:
                    ed_video = len(input_tokens) + 1
            else:
                ed_video = len(input_tokens) + 1
            if ed_image < ed_video:
                t, h, w = (
                    image_grid_thw[image_index][0],
                    image_grid_thw[image_index][1],
                    image_grid_thw[image_index][2],
                )
                image_index += 1
                remain_images -= 1
                ed = ed_image
            else:
                t, h, w = (
                    video_grid_thw[video_index][0],
                    video_grid_thw[video_index][1],
                    video_grid_thw[video_index][2],
                )
                video_second_per_grid_t = 1.0
                if second_per_grid_ts:
                    video_second_per_grid_t = second_per_grid_ts[video_index]
                video_index += 1
                remain_videos -= 1
                ed = ed_video

            llm_grid_t, llm_grid_h, llm_grid_w = \
                t, h // spatial_merge_size, w // spatial_merge_size
            text_len = ed - st

            st_idx = llm_pos_ids_list[-1].max() + 1 if len(
                llm_pos_ids_list) > 0 else 0
            llm_pos_ids_list.append(
                torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)

            t_index = (torch.arange(llm_grid_t).view(-1, 1).expand(
                -1, llm_grid_h * llm_grid_w) * video_second_per_grid_t *
                       tokens_per_second).long().flatten()

            h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(
                llm_grid_t, -1, llm_grid_w).flatten()
            w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(
                llm_grid_t, llm_grid_h, -1).flatten()
            llm_pos_ids_list.append(
                torch.stack([t_index, h_index, w_index]) + text_len + st_idx)
            st = ed + llm_grid_t * llm_grid_h * llm_grid_w

        if st < len(input_tokens):
            st_idx = llm_pos_ids_list[-1].max() + 1 if len(
                llm_pos_ids_list) > 0 else 0
            text_len = len(input_tokens) - st
            llm_pos_ids_list.append(
                torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)

        llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
        mrope_position_delta = (llm_positions.max() + 1 -
                                len(input_tokens)).item()
        llm_positions = llm_positions[:, context_len:seq_len]

        return llm_positions, mrope_position_delta

    @classmethod
    def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
        if modality.startswith("image"):
            return "<|vision_start|><|image_pad|><|vision_end|>"
        if modality.startswith("video"):
            return "<|vision_start|><|video_pad|><|vision_end|>"

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

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config: Qwen2VLConfig = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config

        self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
        self.config = config
        self.multimodal_config = multimodal_config

        if multimodal_config.get_limit_per_prompt("image") or \
            multimodal_config.get_limit_per_prompt("video"):
            self.visual = Qwen2VisionTransformer(
                config.vision_config,
                norm_eps=getattr(config, "rms_norm_eps", 1e-6),
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "visual"),
                use_data_parallel=self.use_data_parallel,
            )
        else:
            self.visual = None

        self.language_model = init_vllm_registered_model(
            vllm_config=vllm_config,
            prefix=maybe_prefix(prefix, "language_model"),
            architectures=["Qwen2ForCausalLM"],
        )

        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}. "
                             f"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[Qwen2VLImageInputs]:
        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 Qwen2VLImagePixelInputs(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 Qwen2VLImageEmbeddingInputs(type="image_embeds",
                                               image_embeds=image_embeds,
                                               image_grid_thw=image_grid_thw)

    def _parse_and_validate_video_input(
            self, **kwargs: object) -> Optional[Qwen2VLVideoInputs]:
        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 Qwen2VLVideoPixelInputs(
                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 Qwen2VLVideoEmbeddingInputs(type="video_embeds",
                                               video_embeds=video_embeds,
                                               video_grid_thw=video_grid_thw)

    def _process_image_input(
            self, image_input: Qwen2VLImageInputs) -> 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"]
        else:
            pixel_values = image_input["pixel_values"]

            if self.use_data_parallel:
                return run_dp_sharded_mrope_vision_model(self.visual,
                                                         pixel_values,
                                                         grid_thw_list,
                                                         rope_type="rope_3d")
            else:
                image_embeds = self.visual(pixel_values,
                                           grid_thw=grid_thw_list)

        # Split concatenated embeddings for each image 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 image_embeds.split(sizes)

    def _process_video_input(
            self, video_input: Qwen2VLVideoInputs) -> 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"]
        else:
            pixel_values_videos = video_input["pixel_values_videos"]
            if self.use_data_parallel:
                return run_dp_sharded_mrope_vision_model(self.visual,
                                                         pixel_values_videos,
                                                         grid_thw_list,
                                                         rope_type="rope_3d")
            else:
                video_embeds = self.visual(pixel_values_videos,
                                           grid_thw=grid_thw_list)

        # 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:
        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",
                             "image_embeds") and "images" not in modalities:
                modalities["images"] = self._parse_and_validate_image_input(
                    **kwargs)
            if input_key in ("pixel_values_videos",
                             "video_embeds") and "videos" not in modalities:
                modalities["videos"] = self._parse_and_validate_video_input(
                    **kwargs)

        return modalities

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

    def get_multimodal_embeddings(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"]
                vision_embeddings = self._process_image_input(image_input)
                multimodal_embeddings += vision_embeddings
            if modality == "videos":
                video_input = modalities["videos"]
                video_embeddings = self._process_video_input(video_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 Qwen2-VL.

        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 Qwen2-VL
                opensource models), the shape will be `(3, seq_len)`,
                otherwise it will be `(seq_len,)`.
            intermediate_tensors: Intermediate tensors from prior forward pass.
            inputs_embeds: Optional tensor of input embeddings.
        """

        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]:

        skip_prefixes = []
        if self.visual is None:
            skip_prefixes.extend(["visual."])
        loader = AutoWeightsLoader(self, skip_prefixes=skip_prefixes)
        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",
            connector="visual.merger.",
            tower_model="visual.",
        )