Skip to content

vllm_omni.model_executor.models.indextts2.codec.quantizer

Decode-only factorized residual vector quantizer.

Adapted from the Amphion codec implementation vendored by IndexTTS 2.5. Training-only paths are intentionally omitted.

FactorizedVectorQuantizeDecoder

Bases: Module

codebook instance-attribute 

codebook = nn.Embedding(codebook_size, codebook_dim)

codebook_dim instance-attribute 

codebook_dim = codebook_dim

in_project instance-attribute 

in_project = _weight_norm_conv1d(
    input_dim, codebook_dim, kernel_size=1
)

input_dim instance-attribute 

input_dim = input_dim

out_project instance-attribute 

out_project = _weight_norm_conv1d(
    codebook_dim, input_dim, kernel_size=1
)

vq2emb 

vq2emb(codes: Tensor) -> Tensor

ResidualVQDecoder

Bases: Module

num_quantizers instance-attribute 

num_quantizers = num_quantizers

quantizers instance-attribute 

quantizers = nn.ModuleList(
    [
        FactorizedVectorQuantizeDecoder(
            input_dim=input_dim,
            codebook_size=codebook_size,
            codebook_dim=codebook_dim,
        )
        for _ in range(num_quantizers)
    ]
)

vq2emb 

vq2emb(
    codes: Tensor, n_quantizers: int | None = None
) -> Tensor