q: [B, Lq, Nq, C1]. k: [B, Lk, Nk, C1]. v: [B, Lk, Nk, C2]. Nq must be divisible by Nk. q_lens

video: one video with shape [1, C, T, H, W].

x: [B, L, N, C]. grid_sizes: [B, 3]. freqs: [M, C // 2].

r""" Reconstruct video tensors from patch embeddings. Args: x (List[Tensor]): List of patchified features

r""" Forward pass through the diffusion model Args: x (List[Tensor]): List of input video tensors, each w

Args: query (torch.tensor): B M H K key (torch.tensor): B M H K shape (tuple): (N_t, N_h, N_w) ref_target_masks: [B, N

PyTorch GPU版本：RGB转Lab颜色空间（输入范围[0,1]，张量形状任意，最后一维为通道数） 参考CIE 1931标准转换公式

r""" Reconstruct video tensors from patch embeddings. Args: x (List[Tensor]): List of patchified features

Autoencoder3d adapted from Stable Diffusion 1.x, 2.x and XL.

r""" Args: x(Tensor): Shape [B, L1, C] e(Tensor): Shape [B, C]

r""" Args: x(Tensor): Shape [B, L1, C] e(Tensor): Shape [B, C]

r""" Initialize model parameters using Xavier initialization.

r""" Initialize model parameters using Xavier initialization.

PyTorch GPU版本：Lab转RGB颜色空间（输出范围[0,1]，张量形状任意，最后一维为通道数）

全GPU批量运算版本：将视频chunk的颜色匹配到参考图像并混合（支持B>1、T帧并行） Args: source_chunk (torch.Tensor): 视频chunk (B, C, T, H, W)，范围[-1, 1] refere

Resize image or tensor to the target size without padding.

Executes the inference using torchrun for Multi-GPU support.

Helper function to save the video, similar to generate_video.py but adapted for function usage.