Apply chunk-causal softmax attention to a token sequence.

Forward pass through the RMSNorm layer. Args: x (torch.Tensor): The input tensor. Returns:

Apply RMS normalization along the configured dimension.

Forward pass with optional causal convolution and feature caching. Args: hidden_states: Input tensor [B, C, T, H, W]

Forward pass through ResNet block.

Forward pass with temporal grouping state management.

Forward pass with trim state management.

Full forward pass: encode then decode. Args: sample: Input video tensor [B, C, T, H, W] temb: Optional timestep embed

forward pass, compatible with WanDiffusionWrapper interface

dpm solver donnot need variance prediction

dpm solver donnot need variance prediction

dpm solver donnot need variance prediction

dpm solver donnot need variance prediction

dpm solver donnot need variance prediction

dpm solver donnot need variance prediction

Bidi state-cached chunkwise GDN with shared Phase A and combined-history Phase B. Default chunkwise path for ``_fused_statecached_forward``.

Drop-in replacement for `fused_gdn.fused_bigdn_stateful` using the chunkwise pipeline. Same signature, same return shape: output (B, N, H, D

Single-direction chunkwise GDN — drop-in for `fused_gdn.fused_gdn_func`. Computes only one scan direction (Phase B + Phase C × 1) and returns

Single-direction chunkwise GDN with optional state cache — drop-in for `fused_gdn.fused_gdn_stateful`. Forward direction supports state load/save