r""" Returns the default logging level.

r""" Returns the jinja template.

r""" Gets paligemma token type ids for computing loss. Returns: batch_token_type_ids: shape (batch_size, sequence_length)

r""" Prepares the tokenized dataset to perform AutoGPTQ. Do not use tensor output for JSON serialization.

r""" Loads a single dataset and aligns it to the standard format.

r""" Pads the tensor to the same length as the target tensor.

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

r""" Aborts the processes recursively in a bottom-up way.

r""" Sets module as a leaf module to skip partitioning in deepspeed zero3.

r""" Aligned dataset: _prompt: [{"role": "user", "content": "..."}] * (2T - 1) _response: [{"role": "assistant", "content": "..."}

r""" Casts projector output to half precision for fine-tuning quantized VLMs.

r""" Gets a list of responses of the chat model.

r""" Checks the version of the required packages.

Clean the units in the number.

r""" Computes loss for preference learning.

r""" Computes log probabilities of the reference model.

r""" Computes log probabilities of the reference model.

r""" Priority: PTQ-quantized (train/infer) > AutoGPTQ (export) > On-the-fly quantization (train/infer)

r""" Patches VLMs before loading them.

r""" Creates deepspeed config.

r""" Creates reward model for PPO training.

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

r""" Returns multiple pairs of token ids representing prompts and responses respectively.