r""" Registers a chat template. To add the following chat template: ``` [HUMAN]: user prompt here [AI]: model response he

r""" Returns a logger with the specified name. It it not supposed to be accessed externally.

r""" Gets element by id. Example: top.lang, train.dataset

r""" Forms a list of slots according to the inputs to encode.

r""" Loads pretrained tokenizer and optionally loads processor. Note: including inplace operation of model_args.

r""" Gets chat template and fixes the tokenizer.

r""" Gets the path to saved model checkpoints.

r""" Validates if this model accepts the input modalities.

r""" Loads pretrained model.

r""" Processes visual inputs. Returns: (llava and paligemma) pixel_values: tensor with shape (B, C, H, W) Return

r""" Returns a single pair of token ids representing prompt and response respectively.

r""" Gets the train dataset and optionally gets the evaluation dataset.

r""" Pre-processes input messages before tokenization for VLMs.

r""" Pre-processes token ids after tokenization for VLMs.

r""" Adds elements to manager.

r""" Gets the current available device.

r""" Saves model checkpoint. Subclass and override to inject custom behavior.

r""" Computes the real sequence length after truncation by the cutoff_len.

r""" Casts a torch tensor or a numpy array to a numpy array.

r""" Plots loss curves and saves the image.

r""" Collects GPU or NPU memory.

r""" Loads user config if exists.

Checks template for both the slow tokenizer and the fast tokenizer. Args: model_id: the model id on hugging face hub. templa

r""" Gets the number of available GPU or NPU devices.

r""" Returns the list of all elements.

r""" Builds batched multimodal inputs for VLMs.

r""" Gets the response token-by-token of the chat model.

r""" Returns a list of names of parameters with weight decay. (weights in non-layernorm layers)

r""" Gets arguments to load config/tokenizer/model. Note: including inplace operation of model_args.

r""" Gets the dict to update the components.

r""" Creates reference model for PPO/DPO training. Evaluation mode is not supported. The valuehead parameter is randomly initialized since it

r""" The model is already unwrapped. There are three cases: 1. full tuning without ds_zero3: state_dict = {"model.layers.*": ..., "v_head

r""" Converts dataset examples to messages.

r""" Gets the base elements that are commonly used.

r""" Generates model's responses given queries.

r""" Gets logits processor that removes NaN and Inf logits.

r""" Loads model config.

Normalizes answers by removing articles, punctuation, fixing whitespace, and optionally converting to lowercase.

Normalizes answers by removing articles, punctuation, fixing whitespace, and optionally converting to lowercase.

r""" Converts elements to token ids.

r""" Encodes formatted inputs to pairs of token ids. Turn 0: prefix + system + query resp Turn t: sep + query

r""" Returns the jinja template.