Get the full file path for lazy loading, handling both file and directory inputs. Args: lazy_load_file: Base file or directory path for l

Get the local rank, the local index of the GPU.

Run inference with action and camera pose conditioning. Args: inputs: Dict containing: - text_encoder_ou

Load metadata from path

Move base attributes and LoRA attributes to CPU device. Args: cls: Target class instance containing tensor attributes base_attrs:

Move base attributes and LoRA attributes to CUDA device. Args: cls: Target class instance containing tensor attributes base_attrs

Produce one config into ring buffer and advance tail by rdma_faa.

Register a self-attention KV cache pool implementation.

Use port active MTU for RTR path_mtu (avoids hard-coded 1024 vs link mismatch).

Run text encoder Args: prompt: Input text prompt img: Optional input image (for some models) Returns:

Load configuration for Wan model.

Set timesteps for the scheduler.

Sets the discrete timesteps used for the diffusion chain (to be run before inference). Args: num_inference_steps (`int`):

Applies the swish activation function.

Avoid leaking reference-video latents into non-v2av runs on a reused runner, and avoid re-appending reference tokens in stage-2 upsampling.

Estimate upper-bound byte sizes for each RDMA buffer slot.

Get input/output tensor names from engine.

Load a TensorRT engine from file.

Return logical->physical chunks as ``(logical_start, physical_start, length)``. Chunks are returned in logical order. They address the physic