Check whether a file path exists. Args: filepath (str or Path): Path to be checked whether exists. Returns:

Generate samples from the batch. Based on given batch, it will automatically determine whether to generate image or video samples. Ar

Generate a streaming video autoregressively with KV cache. Args: stateless_kv: If True, use stateless KV cache mode where all cac

Get unit-norm camera rays in world coordinates for each pixel center. Args: cam_pose (torch.Tensor/np.ndarray [...,3,4]): Camera

Assumes a video2world model that also supports text2world mode.

Get the index of the trajectory in the dataset by the trajectory ID. This is useful when you need to get the trajectory length or sampling wei

Initializes a device mesh for use with Hybrid Sharding strategy in FSDP (HSDP) training. This function requires explicit sizes for replica

Invert a camera pose. Args: cam_pose (torch.Tensor/np.ndarray [...,3,4]): Input camera pose (world2cam [R|t]). Returns:

We hanlde two types of data_batch. One comes from a joint_dataloader where "dataset_name" can be used to differenciate image_batch and video_batch.

Check whether a file path is a directory. Args: filepath (str or Path): Path to be checked whether it is a direct

Check whether a file path is a directory. Args: filepath (str or Path): Path to be checked whether it is a direct

r"""Function for padding and resizing a numpy array. Args: arr (np.ndarray): Input array ntokens (int): Number of output tokens af

parse a command string into an api command (e.g. text2image) and a dictionary of args

Write bytes to a given ``filepath``. Args: obj (bytes): Data to be saved. filepath (str or Path): Path to write data.

Write bytes to a given ``filepath``. Args: obj (bytes): Data to be saved. filepath (str or Path): Path to write data.

Perform a regularized Euler step based on x0 prediction. Args: x_s: Current state tensor. s: Current time tensor. t:

r""" Resizes and crops the input video tensor while preserving aspect ratio. The video is first resized so that the smaller dimension matches

Recursively delete a directory tree. Args: dir_path (str or Path): A directory to be removed. Examples: >>>

Rotates half the hidden dims of the input.

Start the watchdog monitoring thread. If the thread is already running, this method does nothing.

Validate on the full validation dataset. Args: model (ImaginaireModel): The PyTorch model. dataloader_val (torch.util

Context manager for monitoring an operation. This is the primary interface for using the watchdog. It automatically tracks the start

Given the camera pose, transform input 3D points from world coordinates to camera coordinates. Args: points (torch.Tensor/np.ndar

r"""Base augmentor class Args: input_keys (list): List of input keys output_keys (list): List of output keys

Waits for a job to complete. Returns (tuple): - job_status: The status of the job - result: The result of the job -

Check if the background thread has encountered an error and raise it if so.

Prepares the input data batch for the diffusion model. Constructs a dictionary containing the video tensor, text embeddings,

Return a file backend based on the prefix or backend_args. Args: prefix (str): Prefix of uri. backend_args (dict): Arguments to i

Translate an object key to an S3 URL for easy_io. Args: key (str): The key of the object. Returns: str: The

Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/). Explanation:

Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/). Explanation:

Broadcast and split the input data and condition for model parallelism. Currently, we only support context parallelism. Comp

Broadcast and split the input data and condition for model parallelism. Currently, we only support context parallelism.

Broadcast and split the input data and condition for model parallelism. Currently, we only support context parallelism.

r"""Function for building data augmentors from augmentor config.

Chooses an FMHA backend in NATTEN (cutlass-fmha, hopper-fmha, blackwell-fmha) for the current use case based on features needed and current G

Clip the gradient norm of an iterable of parameters. Gradient norm clipping requires computing the gradient norm over the entire model.

Compose a sequence of camera transformations together. pose_new = compose_poses([pose_1, pose_2, ... pose_N]) pose_new(x) = pose_N o

Copy a file src to dst and return the destination file. src and dst should have the same prefix. If dst specifies a directory, the fi

Recursively copy an entire directory tree rooted at src to a directory named dst and return the destination directory. src and dst sh

Performs denoising on the input noise data, noise level, and condition Args: xt (torch.Tensor): The input noise data.

Performs denoising on the input noise data, noise level, and condition with interpolation-specific noise handling for conditional fra

Performs denoising on the input noise data, noise level, and condition Args: xt (torch.Tensor): The input noise data.

Helper determining whether dtype is supported with different sets of supported dtypes for training and inference (forward+backward and forwar

Ends time measurements. NOTE: must be done on the same CUDA device and stream as start().

Check whether a file path exists. Args: filepath (str or Path): Path to be checked whether exists. Returns: