This returns the quantizer that models the higher levels of the hierarchy (typically acoustic).

This returns the quantizer that models the higher levels of the hierarchy (typically acoustic). In this case, it's the quantizer itself.

Cardinality of each codebook.

Cardinality of each codebook.

Cardinality of each codebook.

Just calls `CUDAGraphed` on the given function.

Decode the given codes to the quantized representation.

Decode the given codes to the quantized representation.

Decode the given codes to the quantized representation. In the case of the DummyQuantizer, the codes are actually identical to the inp

Converts integer codes into quantized vectors.

Converts the integer codes into quantized vectors.

See `MimiModel.decode`.

Decode from the discrete codes to continuous latent space.

Whether to apply ema to the codebooks.

Encode a given input tensor with the specified frame rate at the given bandwidth. The RVQ encode method sets the appropriate number of quantiz

Encode a given input tensor with the specified sample rate at the given bandwidth.

Encode a given input tensor with the specified sample rate at the given bandwidth. In the case of the DummyQuantizer, the codes are actually i

Encodes `x` into discrete integer codes.

Encodes `x` into discrete integer codes. If `n_q` is provided, only uses the first `n_q` codebook levels.

See `MimiModel.encode`.

Projects a batch of waveforms to latent space. Args: x (torch.Tensor): Float tensor of shape [B, C, T]. Returns:

Token id for optionally marking the last padding step for a word.

Apply rope rotation to query or key tensor.

Args: x (torch.Tensor): Input tensor of shape [B, C, T] with `C` number of channels. frame_rate (int): frame rate of

Args: x (torch.Tensor): Input tensor of shape [B, C, T] with `C` number of channels. frame_rate (int): frame rate of

Given input tensor x, returns first the quantized (or approximately quantized) representation along with quantized codes, bandwidth,

Args: x (torch.Tensor): input tensor to quantize, of shape `[B, C, T]`. n_q (int or None): if provided, number of cod

Return a pretrained Mimi model.

Return a pretrained Moshi LM model. Args: filename: Path to model weights. copy_missing_weights: Whether to copy missing weights

Token id for the start of sequence (audio).