r""" Prepares the indices and seqlens for flash attn varlen function. Returns: indices: indices of non-masked tokens from the flatten

r""" Gets the peft model for the pretrained model with unsloth. Used in training.

r""" Get the vision_feature_select_strategy.

r""" Judges if the model is a vision language model.

r""" An efficient greedy algorithm with binary search for the knapsack problem.

r""" Checks if the path has a tokenized dataset.

r""" Infers the optimal dtype according to the model_dtype and device compatibility.

r""" Initializes the adapters. Support full-parameter, freeze and LoRA training. Note that the trainable parameters must be cast to floa

r""" Checks if the GPU or NPU is available.

r""" Loads dataset_info.json.

r""" Loads peft model with unsloth. Used in both training and inference.

r""" Optionally loads pretrained model with unsloth. Used in training.

Converts all characters in the text to lowercase.

Converts all characters in the text to lowercase.

r""" Merges multiple datasets to a unified dataset.

r""" Computes ORPO's odds ratio (OR) loss for batched log probabilities of the policy model.

r""" Freezes vision tower for VLM LoRA tuning.

r""" Implements training loop for the PPO stage, like _inner_training_loop() in Huggingface's Trainer.

r""" Includes: (1) cast the layernorm in fp32 (2) make output embedding layer require grads (3) add the upcasting of the l

Removes articles ('a', 'an', 'the') from the text.

Removes articles ('a', 'an', 'the') from the text.

r""" Resize token embeddings.

r""" Restores the layernorm parameters in the model. The model is already unwrapped (and gathered).

r""" Saves arguments to launch training.

r""" Saves model predictions to `output_dir`. A custom behavior that not contained in Seq2SeqTrainer.

r""" Saves model predictions to `output_dir`. A custom behavior that not contained in Seq2SeqTrainer.

r""" Finds the index of largest number that fits into the knapsack with the given capacity.

r""" Computes SimPO loss for batched log probabilities of the policy model.

r""" Avoids flash attention import error in custom model files.

r""" Splits the dataset and returns a dataset dict containing train set and validation set. Supports both map dataset and iterable dataset.

r""" Extracts all the function calls from the response message.

r""" Generates the system message describing all the available tools.

r""" Pads batched data to the longest sequence in the batch. We generate 2 * n examples where the first n examples represent chosen e

r""" Pads batched data to the longest sequence in the batch. We generate 2 * n examples where the first n examples represent chosen e

r""" Initializes an inference engine.