Check if the stop signal is received

Write current layer's K/V into the buffer tail and return full K/V views. Args: layer_idx: decoder layer index key_st

通用的模型下载处理函数 Args: model_path_val: 模型路径 model_name: 模型名称（可能包含状态标识） model_category: 模型类别 download_source_val: 下载源 (

检查模型是否已下载 支持子目录路径，如 Qwen-Image-2512/qwen_image_2512_fp8_e4m3fn_scaled.safetensors

格式化模型选项，添加下载状态标识

Get sequence parallel process group.

Get sequence parallel world size.

从缓存的模型列表中获取模型（不再实时从 HF 获取）

Allgather tensors with potentially variable chunk sizes across ranks. Pads each chunk to max_chunk_size, allgathers, then extracts valid fram

Match nn.Linear under torch.amp.autocast(param_dtype): fp32 input + bf16 weight -> bf16 matmul.

Fused residual update for Wan FFN. A 1D gate uses the CUTLASS epilogue contract and applies gate/residual to the GEMM accumulator before the

Temporarily unlock in context manager form, restoring original state on exit. Typical usage: with d.temporarily_unlocked(

Unified attention inference method supporting both TP and non-TP modes. Args: attn_phase: LTX2Attention or LTX2Attention

videos: A list of videos each with shape [C, T, H, W].

Load and build the text encoder model. Returns: GemmaTextEncoder: The fully initialized text encoder model

扫描 model_path 目录，返回可用的文件和子目录

f""" ADD NOISE: x_t = (1 - sigma_t) * x_0 + sigma_t * N ------ self.add_noise(...) => x_t = sigma_t * (N - x_0) + x_

AI is creating summary for apply_gate Args: x (torch.Tensor): input tensor. gate (torch.Tensor, optional): gate tensor. Defaults

Forward with optional gradient checkpointing.

Get sequence parallel rank.

Initialize a Causal-3D convolution layer. Parameters: inflation_mode: Listed as below. It's compatible with all the 3D-VAE checkpoint

Save ComfyUI Image tensor to video file Args: images: ComfyUI Image tensor [N, H, W, C] in range [0, 1] output_path: Path to

Create sinusoidal timestep embeddings. :param t: a 1-D Tensor of N indices, one per batch element. These ma