This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, he

Round 'number' to the nearest integer multiple of 'factor'.

Set the verbosity level for the 🤗 M4's root logger. Args: verbosity (`int`): Logging level, e.g., one of: -

Rescale (height, width) so that: - aspect ratio <= max_ratio, - total area in [min_pixels, max_pixels], - each dimension is mul

Update the summary chat display with new message

returns the model if supported, asserts otherwise

Splits hidden_size dim into attn_head_size and num_heads

Applies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch.Tensor`): T

Check if the given string is a number

Args texts: List[str] images: (:obj:`PIL.Image.Image`, :obj:`np.ndarray`, :obj:`torch.Tensor`): The image to be p

`split` is "train", "validation" or "test".

`split` is "train", "validation" or "test".

Floor 'number' to the nearest integer multiple of 'factor'.

Adapted from datasets.

Loads the state_dict at `m4_states_{process_index}.pkl`

returns config_class, model_class for a given model name or path

Saves the model. Supports big models or PEFT seamlessly if not using DeepSpeed.

Saves state_dict to `m4_states_{process_index}.pkl`

Set the verbosity to the `INFO` level.

This method also removes concatenated spaces.

There are different whitespace characters.

If possible, modifies `media_info`to add clip scores on available texts

Convert a 1D integer-coded mask (like [1,1,1,2,2,2,2,3,3,3,0,0,...]) into sub-sequence lengths. We assume sub-sequence IDs appear in ascendin