Apply modulation to input tensor

Collapse different naming conventions into the runner's canonical keys.

简单的轮询

Prepare multitask mask training logic. Args: task_type: Type of task ("i2v" or "t2v") cond_latents: Conditio

Prepare positional embeddings.

Run TensorRT inference.

Convert numpy/list/scalar inputs into CPU tensors for normalization.

Whether this Wan variant keeps a fixed latent prefix during diffusion. Vanilla Wan2.2 i2v/s2v already does this for the first encoded frames.

Process network output to extract attribute (e.g. points, depth, etc.) and confidence values. Args: out_head: Network ou

将输入张量从 [seq_len, heads/N, hidden_dims] 转换为 [seq_len/N, heads, hidden_dims] 的格式。 参数: input (torch.Tensor): 输入张量，形状为 [seq_len, heads/N

Apply the configured distributed parallel strategy exactly once.

Batched KMeans clustering in PyTorch using Euclidean distance. Args: x: Tensor of shape (B, N, D), batch_size B, N points per batch,

Build the full names of LoRA (down/up/alpha) and weight difference tensors. Args: weight_name: Original weight tensor name lora_p

Returns the smallest integer greater than or equal to 'number' that is divisible by 'factor'.

Create one service instance on an idle GPU and add it to scheduling pool.

Draw keypoints and connections representing hand pose on a given canvas. Args: canvas (np.ndarray): A 3D numpy array representing th

Encode images using the vision encoder. Args: images: Input images (numpy array or preprocessed tensor) Returns

FP8 GEMM compatible with sgl_kernel.fp8_scaled_mm

获取 Diffusion 模型可选项

Combine encoder outputs for i2v task Args: clip_encoder_out: CLIP encoder output vae_encoder_out: VAE encoder output

获取高噪模型可选项

Return the temporal bounds [inclusive start, exclusive end) for every patch. Returns shape [1, time_steps, 2] where 1 is temporal dim

Get the world size, the total amount of GPUs.

Run a transformer block with kv cache.

Load audio as ``(tensor, sample_rate)``, matching ``torchaudio.load`` defaults.

负载均衡all-to-all通信实现

加载safetensors文件或目录，支持按key包含筛选或排除

Load state dictionary into class instance, resolving block indices for base and LoRA attributes. Args: cls: Target class instance to load

Load text encoder based on text_encoder_type configuration. Supported types: - "lightllm_service": LightLLM HTTP service mode