Copies the `state` . This method copies the data in `state` using `copy.deepcopy()`. If this is not appropriate for your constraint,

Do adaptive gradient clipping on a param group. If ``collect_param_metrics`` is ``True`` this will return the total number of gradie

Do global fixed gradient clipping on a param group. If ``collect_param_metrics`` is ``True`` this will return the total number of gr

Load the state produced from `self._get_flat_param_state_to_save()`.

Computes scaled dot product attention on query, key and value tensors, using an optional attention mask if passed, and applying dropo

Split a full s3 path into the bucket name and path.

This unsharding implementation consists of: 1. Loading each shard on a separate process and copying their sharded tensors to shared

Expand tensors in a state dictionary from `(batch_size, *)` to `(batch_size * beam_size, *)`.

Calculates ANLS for DocVQA. There does not seem to be an official evaluation script. Public implementation on which this implementation is ba

Converts `patch_order` into an array mapping patch_id -> token_position

Check if the given string a number.

Clips gradients for every group that has the field `max_grad_norm`. At the same time collect metrics for each parameter and its gradi

: param images: (batch_size, num_crops, num_patch, n_pixels)

Evaluate a multiple choice instance.

Evaluate an open question instance

Build HTML visualizations for an examples ex: The example (after preprocessing) to show preprocessor: The preprocessor used to preprocessor t

Exact all forms of numbers from a string with regex.

Make sure old optim state dicts are compatible with new versions.

Applies system formatting to ith message from a sequence of messages