Validates multi-dimensional parameters.

Converts all multi-dimensional parameters to standard types.

r"""Function for obtaining the image size from the data dict. Args: data_dict (dict): Input data dict input_keys (list): List of

update the net_ema

The model preparation before the training is launched Args: memory_format (torch.memory_format): Memory format of the model.

Called at the end of a training step for every batch even when using gradient accumulation. This is paired with on_training_step_batc

Called before the training step, for each batch, similarly to on_training_step_start(). This function is paired with on_training_step

2D context padding: simultaneously perform padding on both height and width dimensions

Convert optional path to optional string.

Compute the first order phi function: (exp(t) - 1) / t. Args: t: Input tensor. Returns: Tensor: Result of phi1 function

r"""Creates a post-merging function after all-to-all communication. Args: local_seq_2_local_head (bool): If True, moves sequence dimension

Run the postprocessing on the video frames.

Write data to a given ``filepath`` with 'w' mode. Note: ``put_text`` should create a directory if the directory of ``

Read a video file and extract its frames and metadata.

Remove a file. Args: filepath (str or Path): Path to be removed. Raises: FileNotFoundError: If filepath does

r"""Function to repeat the list to a fixed shape. n is the desired length of the extended list. Args: x (list): Input list n (

Replace the self-attention operator with a sparse self-attention operator. Args: model: MiniTrainDIT instance n_dense_blocks

Reports measurements.

Perform a residual-based 2nd order Runge-Kutta step. Args: x_s: Current state tensor. t: Target time tensor. s: Curr

Resets recorded measurements

Reset/initialize the KV caches. Only has effect in stateful mode. In stateless mode this is a no-op. Args: max_cache_si

Resize the image to the given size. This is done so that wandb can display the image correctly.

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

Perform a robust broadcast operation that works regardless of tensor shapes on different ranks. The function is decorated with @torch.compil

Optimized version of rope_apply using flash_attention's rotary embedding implementation. This version processes the entire batch at once for

Run VQA inference on a video with multiple questions from a YAML file. Args: video_path: Path to the video file

r"""This method calls the `TrainTimeSampler` to sample training times. Returns: t (`torch.Tensor`): A tensor of s

Save network weights, optimizer parameters, scheduler parameters to a checkpoint. Args: model (ImaginaireModel): The PyTorch mode

Save network weights, optimizer parameters, scheduler parameters to a checkpoint. Args: model (ImaginaireModel): The PyTorch mode

Save network weights, optimizer parameters, scheduler parameters to a checkpoint. Args: model (ImaginaireModel): The PyTorch mode

Saves a Config object to a file using pickle serialization. This method is typically used when the configuration object contains comp

Set the metadata for the transforms. This is useful for transforms that need to know the metadata, such as the normalization values.

Convert sRGB values to physically linear ones. The transformation is uniform in RGB, so *srgb* can be of any shape. *srgb* values should rang

Runs a single inference step to generate the next video frame and the full video given an input image and action. Args:

Gracefully shutdown all worker processes. Performs orderly shutdown by: 1. Broadcasting shutdown command to all workers 2. Te

Modify based on DDP source code Synchronizes the parameters and buffers of a model across different processes in a distributed setting.

The training function. Args: model (ImaginaireModel): The PyTorch model. dataloader_train (torch.utils.data.DataLoade

Performs a single training step for the diffusion model. This method is responsible for executing one iteration of the model's train

Update the current frame. Frame should be HWC, uint8, RGB.

Run a forward pass with ``store_kv=True`` to populate the KV cache. In stateful mode (``kv_states=None``), the cache is updated in-place on

Validate if the actual answer contains the expected answer or at least one of the expected keywords. Args: actual_answer

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

Validate if the checkpoint_id is a valid S3 URI.