Normalizes video data in-place on a CUDA device to reduce data loading overhead. This function modifies the video data tensor within

r"""Function for obtaining text embeddings and text mask. Args: cfg (dict): Config dict embeddings_captions (np.ndarray): Caption

Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape `(batch_size, key_value_l

Register a backend. Args: name (str): The name of the registered backend. backend (BaseStorageBackend): The backend class to be r

Old version of the model is stored with all names of parameters sharing common prefix 'module.

Reshapes a batch of images for visualization, organizing them into a grid format. Args: img_data (torch.Tensor | np.ndarray): The im

Run a single chunk of generation. Returns (latents, video). Args: cond_frames: Conditioning frames, shape ``[B, C, T, H, W]`` (ui

"Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the prediction of x_0 at t

Create a transformation pipeline for preprocessing images. Args: n_px (int): The desired size of the transformed image. Returns

Add LoRA (Low-Rank Adaptation) adapters to `self.net`. This function injects LoRA adapters into specified modules of the network, ena

Gather the corresponding tensor from all GPU devices to a list. Args: tensor (torch.Tensor): Pytorch tensor. Returns: tensor

Performs the backward (denoising) process with the student net to get the noisy examples x_t'. See Sec. 4.5 of https://arxiv.org/pdf/

Broadcast non-blocking a task to all workers.

Broadcast a string to all ranks.

r""" Build the dataset object. The function only diffs from BaseDataset.build_dataset by only adding the decoder_handler to the WebDat

Wrap one optimizer per model part in an OptimizersContainer which provides a single step() and zero_grad() method for all the child optimizers.

Cache text encoder checkpoint from S3 to local disk. Args: s3_ckpt_path: S3 path to the text encoder checkpoint cache_dir: L

Calculate the dataset statistics of all columns for a list of parquet files.

Calculate the start and end indices for a given rank in a distributed setting. Args: dataset_length (int): The total length of the d

Censor the prompt using the blocklist with better-profanity fuzzy matching. Args: input_prompt: input prompt to censor R

Chooses an FNA backend in NATTEN (cutlass-fna, hopper-fna, blackwell-fna) for the current use case based on features needed and current GPU a

Clean up distributed resources.

Clean up environment.

Composite the samples to render the corresponding pixels. Args: quantities (torch.Tensor [...,samples,channels]): The quantity to compute

Checks if the specified object exists in the S3 bucket with retry logic for errors. Args: filepath (str): The s3 path of

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

Create a fresh empty KV cache state for a single layer. Returns a ``(k_cache, v_cache)`` tuple of None-filled lists, suitable for pas

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

Args: xt (torch.Tensor): The input noise data. sigma (torch.Tensor): The noise level. condition (Text2Wor

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

Args: xt (torch.Tensor): The input noise data. sigma (torch.Tensor): The noise level. condition (Text2Wor

Denoise one step with EDM preconditioning. When ``kv_states`` is not None (stateless KV cache mode), returns ``(x0_pred, new_kv_state

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

This function is used when self.config.use_flowunipc_scheduler is set.

Display current frame. Returns False if window is closed. Call this in main loop.

Wrap the distributed initialization, used for telemetry and timers

Calculate the beta value for EMA update. weights = weights * beta + (1 - beta) * new_weights Args: iteration (in

Euler angle (ZYX) to rotation matrix.

Evaluate the success rate from a VQA results JSON file. Args: output_path: Path to the JSON output file threshold: Success t

Ensure CR1 embeddings are available at the specified path. For INTERNAL builds, verifies the file exists and raises an error if not. For

Convert world2cam extrinsic matrix/matrices [R|t] to (x,y,z,w, tx,ty,tz). Args: cam_pose (torch.Tensor/np.ndarray [...,3,4]): Wor

Gradio component that allows users to upload files, browse uploads, and view file contents. Args: upload_dir (str): The directory to

Finalize the checkpointer.

Finalize the checkpointer.

Format time in nanoseconds based on desired unit.

Automatically format time in nanoseconds either automatically or based on desired unit.

r""" Args: x(Tensor): Shape [B, L1, C] e(Tensor): Shape [B, C]