Precise number of tokens that should be sampled at each level. This is mostly useful for running dummy
                experiments
            offset (`int`, *optional*, defaults to 0):
                Audio offset used as conditioning, corresponds to the starting sample in the music. If the offset is
                greater than 0, the lyrics will be shifted take that intoaccount
            save_results (`bool`, *optional*, defaults to `True`):
                Whether or not to save the intermediate results. If `True`, will generate a folder named with the start
                time.
            sample_length (`int`, *optional*):
                Desired length of the generation in samples.

        Returns: torch.Tensor

        Example:

        ```python
        >>> from transformers import AutoTokenizer, JukeboxModel, set_seed
        >>> import torch

        >>> metas = dict(artist="Zac Brown Band", genres="Country", lyrics="I met a traveller from an antique land")
        >>> tokenizer = AutoTokenizer.from_pretrained("openai/jukebox-1b-lyrics")
        >>> model = JukeboxModel.from_pretrained("openai/jukebox-1b-lyrics", min_duration=0).eval()

        >>> labels = tokenizer(**metas)["input_ids"]
        >>> set_seed(0)
        >>> zs = [torch.zeros(1, 0, dtype=torch.long) for _ in range(3)]
        >>> zs = model._sample(zs, labels, [0], sample_length=40 * model.priors[0].raw_to_tokens, save_results=False)
        >>> zs[0]
        tensor([[1853, 1369, 1150, 1869, 1379, 1789,  519,  710, 1306, 1100, 1229,  519,
              353, 1306, 1379, 1053,  519,  653, 1631, 1467, 1229, 1229,   10, 1647,
             1254, 1229, 1306, 1528, 1789,  216, 1631, 1434,  653,  475, 1150, 1528,
             1804,  541, 1804, 1434]])
        ```
        r