inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    Args:
        act_fn (`str`, *optional*, defaults to `"relu"`):
            Activation function of the model.
        nb_discrete_codes (`int`, *optional*, defaults to 2048):
            Number of codes of the VQVAE.
        commit (`float`, *optional*, defaults to 0.02):
            Commit loss multiplier.
        conv_input_shape (`int`, *optional*, defaults to 1):
            Number of audio channels.
        conv_res_scale (`bool`, *optional*, defaults to `False`):
            Whether or not to scale the residuals of the `JukeboxResConv1DBlock`.
        embed_dim (`int`, *optional*, defaults to 64):
            Embedding dimension of the codebook vectors.
        hop_fraction (`List[int]`, *optional*, defaults to `[0.125, 0.5, 0.5]`):
            Fraction of non-intersecting window used when continuing the sampling process.
        levels (`int`, *optional*, defaults to 3):
            Number of hierarchical levels that used in the VQVAE.
        lmu (`float`, *optional*, defaults to 0.99):
            Used in the codebook update, exponential moving average coefficient. For more detail refer to Appendix A.1
            of the original [VQVAE paper](https://arxiv.org/pdf/1711.00937v2.pdf)
        multipliers (`List[int]`, *optional*, defaults to `[2, 1, 1]`):
            Depth and width multipliers used for each level. Used on the `res_conv_width` and `res_conv_depth`
        res_conv_depth (`int`, *optional*, defaults to 4):
            Depth of the encoder and decoder block. If no `multipliers` are used, this is the same for each level.
        res_conv_width (`int`, *optional*, defaults to 32):
            Width of the encoder and decoder block. If no `multipliers` are used, this is the same for each level.
        res_convolution_multiplier (`int`, *optional*, defaults to 1):
            Scaling factor of the hidden dimension used in the `JukeboxResConv1DBlock`.
        res_dilation_cycle (`int`, *optional*):
            Dilation cycle value used in the `JukeboxResnet`. If an int is used, each new Conv1 block will have a depth
            reduced by a power of `res_dilation_cycle`.
        res_dilation_growth_rate (`int`, *optional*, defaults to 3):
            Resnet dilation growth rate used in the VQVAE (dilation_growth_rate ** depth)
        res_downs_t (`List[int]`, *optional*, defaults to `[3, 2, 2]`):
            Downsampling rate for each level of the hierarchical VQ-VAE.
        res_strides_t (`List[int]`, *optional*, defaults to `[2, 2, 2]`):
            Stride used for each level of the hierarchical VQ-VAE.
        sample_length (`int`, *optional*, defaults to 1058304):
            Provides the max input shape of the VQVAE. Is used to compute the input shape of each level.
        init_scale (`float`, *optional*, defaults to 0.2):
            Initialization scale.
        zero_out (`bool`, *optional*, defaults to `False`):
            Whether or not to zero out convolution weights when initializing.
    Z
jukebox_vqvaeZ