Propagate the input points across frames to track in the entire video.

Read a local video using imageio. Returns: frames: list of (H, W, 3) uint8 RGB frames

Removes small disconnected regions and holes in a mask. Returns the mask and an indicator of if the mask has been modified.

This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, he

Compute a binary mask from an uncompressed RLE.

Rotates half the hidden dims of the input.

Create directory if it does not exist.

Save a mask as a PNG file with the given palette.

Split a list into N contiguous chunks and return the K-th chunk (1-indexed).

Init. Args: features (int): number of features

Initialize the neck :param trunk: the backbone :param position_encoding: the positional encoding to use :param d_model: the di

Build SAM-style prompt encoder and mask decoder.

dynamic RoPE layers should recompute `inv_freq` in the following situations: 1 - growing beyond the cached sequence length (allow sca

When outputting a single mask, if the stability score from the current single-mask output (based on output token 0) falls below a thr

Embeds box prompts.

Embeds mask inputs.

Embeds point prompts.

尝试在 pipe 里找到“主干网络”，不同版本可能叫： - dit / transformer / unet / model / diffusion_model / backbone 等 找到就返回，否则报错。

Gets the batch size of the output given the batch size of the input prompts.

Get a dummy object pointer based on an empty mask on the current frame.

Get the number of images and videos for each sample to calculate the separation length of the sample tensor. These parameters are not

Compute stability scores of the mask logits based on the IoU between upper and lower thresholds.

Wan2.2 A14B 常见：vace/vace2（以及可能的 dit/dit2） 都做 TP。

WanVideoPipeline 里真正会在 denoise loop 里跑的大块通常是： - dit / dit2 - vace / vace2 有就都做 TP（比只对一个 backbone 更合理）

Apply the RMSNorm normalization to the input tensor. Args: x (torch.Tensor): The input tensor. Returns: