Struct RvqDepthDecoder 

pub struct RvqDepthDecoder { /* private fields */ }

The RVQ depth decoder: given the frame’s hidden state and the codes chosen so far, it predicts the next codebook level.

It is a small CAUSAL transformer over a sequence that never exceeds the codebook count — the positional table is 16 rows for 8 levels — with no rotary embedding at all. Attention that forgets the mask here leaks a level’s own answer backwards and the model still samples.

Implementations

impl RvqDepthDecoder

pub fn from_cmf(model: &Arc<CmfModel>) -> Result<Self, String>

pub fn codebooks(&self) -> usize

pub fn project(&self, x: &[f32]) -> Vec<f32>

projection is applied to every element entering the sequence — the frame hidden, the c0 embedding and each extra embedding.

pub fn forward_last(&self, seq: &[f32], n: usize) -> Vec<f32>

Run the stack over [n, hidden] and return the LAST position’s normed hidden — the only one the caller reads.

pub fn head(&self, level: usize, hidden: &[f32]) -> Vec<f32>

Logits for codebook level (1-based; level 1 uses head 0).