Load from a YAML file.

Load a regular state dict from the file ``fname`` within ``checkpoint_dir`` using :func:`torch.load()`. This can be used during distributed t

Initialize weights of a linear or embedding module. :param config: The model config. :param module: The linear or embedding submodule to

Save to a YAML file.

Given a starting state and a step function, apply beam search to find the most likely target sequences. Returns a tuple of `

Get the local rank per filesystem, meaning that, regardless of the number of nodes, if all ranks share the same filesystem then `get_fs_local_rank

Update the legacy config settings whose schemas have undergone backwards-incompatible changes.

Get a human-readable prompt by compressing the image tokens

Save a regular state dict to the file ``fname`` within ``checkpoint_dir`` using :func:`torch.save()`. This can be used during distributed tra

Download urls from a PixMo dataset, return a map of urls->filename

Builds a HTML table visualization of the predictions

Get the total number of parameters.

Upload source file to a target location on GCS or S3.

Selects a multiple-choice option based on the model output The output is should exactly match one of the option, but contains some heuri

Wait for the condition function to return True.

Groups a pixmo datasets so each row contains all annotation for one image, and add images path using `url_to_path`, removing rows that do not

Get directory to save the eval results

Get the peak GPU memory usage in MB across all ranks. Only rank 0 will get the final result.

Monkeypatch for torch's ShardedTensor, so we can unpickle without having torch.distributed set up.

Turns an image into a string that can be used as a src in html images