Read data from current file managed by storage. Args: output_type: Type that you want to read to, either "datafr

Write data to current file managed by storage. data: Any, the data to write, it should be a dataframe, List[dict], etc.

批量生成文本，支持对话历史格式，与 local_model_vlm_serving 接口对齐 :param conversations: 对话历史列表，每个元素是一个对话历史 格式：[[{"

Generate prompt using custom template.

Load image from various formats (path, bytes, PIL Image). Args: image_data: Image in various formats Return

Generate data from input. input: List[str], the input of the generator

Gets attributes of a node which is polymorphic over return type. From https://github.com/pytorch/pytorch/pull/82628

Recursively copy all contents from source_path to destination_path. Prompts user if a file already exists in destination.

Convert string to float, handling commas.

Collapse whitespace to single spaces and trim.

Convert dataframe to requested output type.

Read an image file and convert it to a base64-encoded string, returning the image data and MIME format. :param image_path: Path to t

Gather python objects across ranks into a list of length `world_size`.

Gather python objects across ranks into a list of length `world_size`.

Gather python objects across ranks into a list of length `world_size`.

Load image file, return PIL Image object (Gemini API can directly accept PIL Image)

去掉_ . -等特殊字符，并转换为小写

Remove caption-related references and replace with video-related terms.

Remove caption-related references and replace with video-related terms.

Get video filename without extension.

这里写一份通用的 run 文档也可以， 不写的话会被下面动态拷贝原 operator.run 的 __doc__。

Extract prompt text from user_input

将格式1的数据转换为格式2。 支持多轮对话和多模态（图片、视频、音频），并验证token数量。

Wrap content items into the standard OpenAI messages structure. :param content: List of content dicts (text/image elements).

将音频文件编码为 base64 字符串 :param audio_path: 音频文件路径 :return: base64 编码的字符串

将图像文件编码为 base64 字符串 :param image_path: 图像文件路径 :return: base64 编码的字符串

将视频文件编码为 base64 字符串 :param video_path: 视频文件路径 :return: base64 编码的字符串

Normalize a cell that may be str or [str] into a plain path string.

Normalize a cell that may be str or [str] into a plain path string.

Normalize a cell that may be str or [str] into a plain path string.

Generate response for a single input with retry logic. Args: user_input: User input text system_prompt: Syste

Get video frame count using OpenCV. method: - "header": use metadata header (fast, may be unreliable in some codecs) - "set"

给 Chat Completions 的 input_audio.format 用的格式字符串 常见：wav / mp3 / m4a / ogg / webm / flac / aac

Write luminance stats back into df[video_clips_key]["clips"][clip_idx]: clip["luminance_mean"], clip["luminance_min"], clip["luminance_ma

Write OCR scores back into df[video_clips_key]["clips"][clip_idx]: clip["ocr_score"]

Write row-level OCR score back into df[output_key].

Write row-level stats back into df columns: - luminance_mean, luminance_min, luminance_max

Write scores back into df[video_clips_key]["clips"][clip_idx]["aesthetic_score"]. This function mutates a copied dataframe and returns it.

从指定文件中加载类。 :param file_path: 脚本文件的路径 :param class_name: 类的名字 :return: 类对象

Load data from local file based on file type.

基础解析逻辑（内部复用）

基础解析逻辑（内部复用）

Parse OpenAI image generation/editing response, return (key, image)

对单条 clip 元信息执行切割。 仅当 row_dict['filtered'] == False 时执行；否则直接跳过。 如果 'filtered' 字段不存在，默认为 False（不过滤）。 返回 (row_list, valid)；valid=True 表示

Process one video path: - detect scenes OR split by fixed interval - trim/split - build output dict with fps & scene entries

Resolve image paths (supports single path or list of paths). If relative paths, resolve them relative to base_dir.

Compute duration in seconds from num_frames and fps if both are valid.