If `evaluation_run` already exists, get the already logged values into a dictionary.

Returns (trainable_params, total_params) across the entire model.

A list, each element is a list of potential captions

Init the wandb logger.

when run under pytest-xdist returns the worker id (int), otherwise returns 0

Return function over iterator that groups key, value pairs into samples.

Check if there are any messages in the current chat

Image splitting strategy. 1) If one side of the original image is larger than `max_image_size`, resize it to `max_image_size` while preservin

Merge text embeddings with image embeddings out-of-place (no in-place indexing). The shapes are something like: - inpu

Merge text embeddings with image embeddings out-of-place (no in-place indexing). The shapes are something like: - inpu

This method aims at merging the token embeddings with the image hidden states into one single sequence of vectors that are fed to the transfo

This method aims at merging the token embeddings with the image hidden states into one single sequence of vectors that are fed to the transfo

Check if the current chat has been modified since loading

Return True if the dataset entry is from MammothVL. We assume the name starts with "mammoth:".

Return True if the dataset entry is from OneVision. We assume the name starts with "onevision:".

Returns the sub-sample weight given the sub-sequence length.

Load a specific chat from disk

Treats a directory of images as if they were consecutive frames in a pseudo-video recorded at `source_fps`, then samples frames to achieve

Treats a directory of images as if they were consecutive frames in a pseudo-video recorded at `source_fps`, then samples frames to achieve

Loads a video from `path` using decord, sampling up to `max_frames` frames. After deduplicating indices (e.g., to handle rounding collisions)

Loads a video from `path` using decord, sampling up to `max_frames` frames. After deduplicating indices (e.g., to handle rounding collisions)