vllm.renderers.hf ¶

def build_video_prompts_from_mm_data(
    mm_data: "MultiModalDataDict",
) -> list[str]:
    """Build video prompts from vision_chunk data.

    Collects prompts from video chunks and groups them by video_idx.

    Args:
        mm_data: Processed multimodal data with vision_chunk items

    Returns:
        List of video prompts, one per video.
    """
    vision_chunks = mm_data.get("vision_chunk")
    if vision_chunks is None:
        return []

    # Group chunks by video_idx
    video_prompts_dict: dict[int, list[str]] = defaultdict(list)

    for item in vision_chunks:
        # vision_chunk items are always dicts (VisionChunkImage/VisionChunkVideo)
        assert isinstance(item, dict)
        if item.get("type") == "video_chunk":
            video_idx = item.get("video_idx", 0)
            prompt = item.get("prompt", "")
            video_prompts_dict[video_idx].append(prompt)

    # Build prompts in video order
    video_prompts = [
        "".join(video_prompts_dict[video_idx])
        for video_idx in sorted(video_prompts_dict.keys())
    ]

    return video_prompts

def rebuild_mm_uuids_from_mm_data(
    mm_uuids: "MultiModalUUIDDict",
    mm_data: "MultiModalDataDict",
) -> "MultiModalUUIDDict":
    """Rebuild mm_uuids after vision_chunk processing.

    When videos are split into chunks, the original UUIDs need to be updated
    to reflect the new UUIDs generated for each chunk.

    Args:
        mm_uuids: Original UUIDs dictionary
        mm_data: Processed multimodal data with vision_chunk items

    Returns:
        Updated UUIDs dictionary with chunk UUIDs
    """
    vision_chunks = mm_data.get("vision_chunk")
    if vision_chunks is None:
        return mm_uuids

    assert all(isinstance(item, dict) for item in vision_chunks), (
        "Expected all vision_chunk items to be dicts"
    )
    vision_chunks = cast(list[dict[str, Any]], vision_chunks)
    vision_chunk_uuids = [
        uuid_val for item in vision_chunks if (uuid_val := item.get("uuid")) is not None
    ]

    if vision_chunk_uuids:
        mm_uuids = dict(mm_uuids)
        mm_uuids["vision_chunk"] = vision_chunk_uuids

    return mm_uuids