Is there a paged_adamw_16bf opim option?

Beginners
1

I saw:

def get_original_training_args(output_dir="./results",
                               per_device_train_batch_size=1,
                               gradient_accumulation_steps=16,  # num its to accumulate before opt update step
                               optim="paged_adamw_32bit",
                               save_steps=10,  # how often to save, if <1 -> % of train steps
                               logging_steps=10,  # how often to log, if <1 -> % of train steps
                               learning_rate=2e-4,
                               max_grad_norm=0.3,
                               max_steps=500,  # number of training steps/its
                               warmup_ratio=0.03,  # number of steps for a linear warmup
                               lr_scheduler_type="constant",
                               ):
    """
    """
    from transformers import TrainingArguments

    training_arguments = TrainingArguments(
        output_dir=output_dir,
        per_device_train_batch_size=per_device_train_batch_size,
        gradient_accumulation_steps=gradient_accumulation_steps,
        optim=optim,
        save_steps=save_steps,
        logging_steps=logging_steps,
        learning_rate=learning_rate,
        fp16=True,
        max_grad_norm=max_grad_norm,
        max_steps=max_steps,
        warmup_ratio=warmup_ratio,
        group_by_length=True,
        lr_scheduler_type=lr_scheduler_type,
    )
    return training_arguments

Is there a 16 brain float option for this? Or should I always have the optimizer be in 32 bit precision?