从多个仓库获取模型并合并

Map an IPv6 GID string to IPv4 when it is IPv4-mapped.

Perform attention for AR model. Note: When running full-video inference (not chunk-by-chunk AR generation), we use bidirecti

Override to handle per-token modulation in post-attention.

Override to handle per-token vec modulation for WorldPlay AR. Returns: Tuple of (img_q, img_k, img_v, img_q_pre_rope, im

Override to handle per-token vec modulation for text branch.

Read tensor metadata (names, shapes, dtypes) from safetensors file header without loading data. Supports two kinds of safetensors files:

Save metadata for old sampling strategy.

Swap a single ``Block``'s attn/norm/mlp leaves with adapters.

Return unsigned circular distance on a 64-bit counter space.

Instantiate a Vocoder from a flat config dict. Args: cfg: Vocoder config dict (keys match Vocoder constructor args). apply_final_a

Apply specified activation function to parameter tensor Args: tensor: Tensor containing parameter values act

Convert diffuser format model to x2v format using converter.py -d backward. Returns path to converted safetensors file.

Wait until the single GPU staging buffer holds ``layer_id``.

A decorator function to assist in performance testing of CUDA operations. This function will: 1. Automatically determine whether any par

Initialization for cache.

Calculate memory usage in bytes for a tensor.

Validate tensor: must be XPU, contiguous, [B, L, H, 128], B=1

Convert checkpoint to safetensors format.