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。

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。

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。

Custom collate function for DataLoader since we have variable-length image lists. We don't stack images; we just return lists.

check whether string is validate base64

Load multiple image files, return PIL Image object list

子进程工作函数

子进程工作函数

静态方法封装，给多进程用

Top-level picklable worker for ProcessPoolExecutor.

发送带有对话历史的单个请求 :param idx: 请求索引 :param conversation: 对话历史 :param image_paths: 图像路径列表 :param video_pat

发送单个请求并返回结果 :param idx: 请求索引 :param prompt: 文本提示 :param image_paths: 图像路径列表 :param video_paths: 视频路径

Batch-tracked version of `analyze_images_with_gpt`, returning (request_id, analysis). :param image_paths: List of image file paths.

输入: messages -> list of list (每个元素是聊天序列) 输出: full_prompts -> list of dict {prompt, multi_modal_data}

输入: messages -> list of list (每个元素是聊天序列) 输出: full_prompts -> list of dict {prompt, multi_modal_data}

输入: messages -> list of list (每个元素是聊天序列) 输出: full_prompts -> list of dict {prompt, multi_modal_data}

参数： - need_fields: 本次会使用到的字段名集合（来自算子的 input_keys.keys()） - kwargs: 每个字段名 -> 该行的值（来自 DataFrame） 逻辑： - 对于每个字段名

Generate system prompt for video quality evaluation.

Generate prompt for CoT QA generation with question parameter.

Same as `chat_with_one_image` but returns a tuple of (request_id, response). :param request_id: Arbitrary identifier for tracking th

Cleanup the generator and garbage collect all GPU/CPU memory.

Clean up any resources (e.g., close HTTP connections).

Clear pipelines and free GPU memory.

Cleanup resources.

Calculate BLEU score.

Copy a single file with path, with checking and asking about the existence.

Copy files under a path without recursion, with checking and asking about the existence.

r""" Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to computing decoding in on

r""" Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to computing decoding in one

Compute EMScore. Args: - :param: `model` : a BERT model in `pytorch_pretrained_bert` - :param: `refs` (list of str): referen

r""" Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to compute decoding in sev

r""" Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and en

[For FunctionalRefiner] 仅返回清洗后的 CoT 文本

[For FunctionalRefiner] 仅返回清洗后的 CoT 文本

[For FunctionalRefiner] 提取并合并关键词

[For FunctionalRefiner] 提取并合并关键词

Args: images_bnc: (B, N, C, H, W) preprocessed images Returns: scores: (B,) averaged aesthetic score per clip

Generate embeddings for a list of texts. Args: texts: List of text strings to embed Returns