Initialize the step_index counter for the scheduler.

Initialize the step_index counter for the scheduler.

Initialize a Per-GPU cross-attention cache for the Wan model.

Initialize a Per-GPU cross-attention cache for the Wan model.

Initialize a Per-GPU KV cache for the Wan model.

Initialize a Per-GPU KV cache for the Wan model.

Inject errors into the clean prefix (E_img). 2D (position-bucketed): the i-th LOCAL prefix block draws from ``buckets[(i, *)]`` with

Inject errors into clean_latent before noising (E_vid). Matches BOTH block_position (LOCAL) and timestep when the buffer is 2D, else

Inject errors into the noise (E_noise). Same matching strategy as ``_inject_latent_error_buffer`` but reads from the dedicated noise

Return True when *chunk_index* is the first chunk of a new scene AND multi-shot sink is enabled.

Load a LoRA checkpoint, unwrapping ``generator_lora`` when present.

Remap ``t_bucket`` to the closest owned timestep bucket.

Mark the current chunk's buffer position as pinned for multi-shot sink. The pinned region REPLACES the original sink on the next rolling even

r""" Prepares and returns the required model for the current timestep. Args: t (torch.Tensor): current ti

r""" Prepares and returns the required model for the current timestep. Args: t (torch.Tensor): current ti

Teacher-Forcing attention mask for the *natural* interleaved layout produced directly by `all_to_all(scatter=head, gather=seq)`:

Restore model attention state saved by _apply_attn_overrides.

On-policy generator via backward simulation (original path).

Off-policy generator: add noise to clean_latent at each timestep in denoising_step_list, randomly pick one, and run the generator for

Sample a random (video_idx, start_frame) valid for the first segment, and from which we can complete ALL segments without wrapping to the star

Override global_sink_size on all submodules; create the attribute if missing.

Override global_sink_size on all submodules; create the attribute if missing.

Set max_attention_size on all submodules that define it. If local_attn_size_value == -1, use the model's global default (32760 for Wa

Set a unified upper bound for all submodules that contain the max_attention_size attribute. local_attn_size_value == -1 indicates glo

Override sink_size on all submodules that define it.

Override sink_size on all submodules that define it.

Initialize the inference pipeline for visualization on CPU, to be moved to GPU only when needed.

Convert complex128 freqs to (cos_f32, sin_f32) — once per cache entry.

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

Generate validation samples to monitor training progress.

Wrap ``pipeline.generator.model`` with a LoRA adapter, load weights, and merge. The merged module ends up structurally identical to the original

Return the 5B VAE wrapper requested by args.vae_type.

Remove old checkpoints if the number exceeds max_checkpoints. Only the main process performs the actual deletion to avoid race condit

Remove old checkpoints if the number exceeds max_checkpoints. Only the main process performs the actual deletion to avoid race condit

Compute the DMD loss (eq 7 in https://arxiv.org/abs/2311.18828). Input: - image_or_video: a tensor with shape [B, F, C, H

Compute DP groups whose ranks each form a Ulysses SP group.

Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is designed to discretize an

r""" Convert the model output to the corresponding type the UniPC algorithm needs. Args: model_output (`torch.Tensor`):

Convert the diffusion network's x0 prediction to noise predidction. x0: the predicted clean data with shape [B, C, H, W] xt: