Compute the value. When current unit is larger than previous unit, current unit * all previous units will be used as all previous uni

Convert applicable model parameters to fp16

一百八 to 一百八十 一亿一千三百万 to 一亿 一千万 三百万

Extract content between <answer> and </answer> tags.

Extract content between <think> and </think> tags.

Build clip metadata per row from prior pipeline columns. Source columns: - `input_video_key`: str or [str] - `video_info_key`: d

Extract frames for each clip in the upstream dataframe. Args: input_video_key: where to read video path(s). video_info_key:

Extract video info for each row in a dataframe. Args: dataframe: Input dataframe (will not be mutated). backend: One of {"op

Compute scene segments for each row. Args: dataframe: input dataframe that already contains video path column, an

对视频的每个clip应用过滤条件，添加过滤标记。 Args: clips: 视频片段列表，每个片段是一个字典，包含质量指标 strict_mode: 是否严格模式，True则缺失列时报错，False则跳过该过滤条件

Format LiteLLM response to include reasoning content in a structured format. This method handles the standardized LiteLLM re

一键跑完当前 pipeline 一个 step： 图片 → Caption

一键生成多模态数学题目数据集： 生成函数图像 + 对应问答对

Distribute inputs across pipelines and run in parallel, then save via image_io.

Process a batch of inputs, return dictionary in {key: [Image, ...]} format. user_inputs format is aligned with LocalImageGen

This method is used for splitting long audios into speech chunks using silero VAD Parameters ---------- audio: torch

Classify the operator type by its path of registration. This is used to classify operators into different categories. :return

判断 serving 是否为 API 类型