r""" Regularizes videos to avoid error. Including reading, resizing and converting.

r""" Asynchronously gets a list of responses of the chat model.

r""" Asynchronously gets a list of scores of the reward model.

r""" Asynchronously gets the response token-by-token of the chat model.

Save actor model on rank 0. Returns: List: list of remote object refs.

r""" Calculates effective tokens per second.

r""" Removes args with NoneType or False or empty string value.

Computes BLEU score of translated segments against one or more references. Args: reference_corpus: list of lists of references for each tra

Run the given model on the given batch of inputs, concatenating the chosen and rejected inputs together. We do this to avoid doing two forwar

r""" Computes the sum log probabilities of the labels under given logits if loss_type is not IPO, ORPO or SimPO. Otherwise the averag

r""" Returns the number of trainable parameters and number of all parameters in the model.

Evaluate predictions using different retrieval docs

The evaluation process after finishing overall generation

r""" Extracts tool message.

r""" Runs forward pass and computes the log probabilities.

r""" Computes the log probabilities of the given labels under the given logits. Returns: logps: A tensor of shape (batch_size,) conta

Used in basic module to get the saved config files.

r""" Gets the path to user config.

r""" Returns an iterator over all elements with their names.

r""" Freezes vision tower and language model for VLM full/freeze tuning.

获取本机IP地址

r""" Builds batched multimodal inputs for VLMs. Arguments: images: a list of image inputs, shape (num_images,)

r""" Returns stop token ids.

Select the maximum number of examplars that can be placed in the prompt

Extract max_tokens and max_new_tokens from a parameter dictionary.

Calculate whether a retrieve is required based on the output probability of the special token in the model

r""" Loads saved arguments.

r""" Loads value head parameters from Hugging Face Hub or local disk. Returns: dict with keys `v_head.summary.weight` and `v_head.summary.bia

Make a list of experience with the micro_rollout_batch_size. This method will first calculate the response sequences and rewards for

r""" Detach `tensors` (even if it's a nested list/tuple/dict of tensors).

Normalize text by removing articles, punctuation, and extra whitespace, and converting to lowercase.

Normalize text by removing articles, punctuation, and extra whitespace, and converting to lowercase.

Parse candidates from model response

Parse model's pair ranking result into score

Parse model's validation result into score based on the paper formula

remote reward model API api_url: RM API, We assume that the API supports two modes: merging query + response and not merging queries: query+re

r""" Replaces the default/reward modules in the model. The model is already unwrapped.

Resolve between max_tokens and max_new_tokens values based on priority. Returns the resolved token value or None if no valid value fo

Perform Reciprocal Rank Fusion (RRF) on retrieval results. Args: results (list of list of dict): Retrieval results for m

r"""Get refining result. Args: item: dataset item, contains question, retrieval result... Returns: str: refi

r""" Saves arguments.

r""" EMA implementation according to TensorBoard.