Struct CodecDecoder 

pub struct CodecDecoder<B: Backend> { /* private fields */ }

Full codec decoder: conv-transformer autoencoder producing 24 kHz audio.

Loads 117 tensors from SafeTensors with weight norm fusion. The decoder_blocks layout is:

• Even blocks (0, 2, 4, 6): convolution (input proj or upsample)
• Odd blocks (1, 3, 5, 7): transformer groups (2 layers each)
• Output: weight-normed Conv1d [1024→240, k=7]

Implementations

impl<B: Backend> CodecDecoder<B>

pub fn from_components( input_conv: CausalConv1d<B>, transformer_group_layers: Vec<Vec<CodecTransformerLayer<B>>>, upsample_convs: Vec<CausalConvTranspose1d<B>>, output_conv: CausalConv1d<B>, vq_codebook: VqCodebook<B>, ) -> Self

Create a codec decoder from pre-loaded components (for GGUF loading).

transformer_group_layers is a Vec of 4 groups, each a Vec of layers.

pub fn from_safetensors( safetensors: &SafeTensors<'_>, config: &CodecDecoderConfig, device: &B::Device, ) -> Result<Self>

Load the full codec decoder from SafeTensors.

Arguments

• safetensors - SafeTensors data containing all codec weights
• config - Codec decoder configuration
• device - Device for tensor allocation

pub fn decode( &self, semantic_indices: &[usize], acoustic_indices: Tensor<B, 2>, ) -> Tensor<B, 2>

Decode quantized tokens into a 24 kHz waveform.

Arguments

• semantic_indices - Semantic token indices per frame, each in [0, 8191]. Shape: [N]
• acoustic_indices - Acoustic FSQ indices per frame [N, 36], each in [0, 20] as f32

Returns

Audio samples [1, total_samples] at 24 kHz.

pub fn vq_codebook(&self) -> &VqCodebook<B>

Access the VQ codebook (for external use).