Save per-block prompts alongside the video. Consecutive identical prompts are merged, e.g.: [0] a, [1] a, [2] b => [0,1] a\\n[2] b\\n

Save prompts for one generated video to a txt file. Consecutive identical prompts are merged, e.g.: [0] a, [1] a, [2] b => [0,1] a\

Return the dequant denominator limits used by FourOverSix ScaleRule.

Set the DP group whose ranks own the same sequence chunk.

Set the SP group used by SP rank/world-size and all-to-all helpers.

Determine whether module should be replaced.

Start the evaluation coordinator.

r""" Generates video frames from text prompt using diffusion process. Args: input_prompt (`str`): Text pr

Input: - timestep: the timestep with shape [B*T] Output: the corresponding weighting [B*T]

Convert a generated video tensor from [T, C, H, W] or [1, T, C, H, W] to uint8 THWC.

Base class for denoising loss. Input: - x: the clean data with shape [B, F, C, H, W] - x_pred: the predicted

r""" Initializes the image-to-video generation model components. Args: config (EasyDict): Object containi

r""" Initializes the Wan text-to-video generation model components. Args: config (EasyDict): Object conta

r""" Initializes the Wan text-to-video generation model components. Args: config (EasyDict): Object conta

r""" Initialize the diffusion model backbone. Args: model_type (`str`, *optional*, defaults to 't2v'): Mo

r""" Initialize the diffusion model backbone. Args: model_type (`str`, *optional*, defaults to 't2v'): Mo

Args: api_key: The API key for Dash Scope authentication and access to related services. model_name: Model name, 'qwe

Args: model_name: Use predefined model names such as 'QwenVL2.5_7B' and 'Qwen2.5_14B', which are specific version