Struct CodecFrameFusedEngine 

pub struct CodecFrameFusedEngine { /* private fields */ }

Warmed codec-frame engine: host CP greedy + optional megagraph cache (parity/bench).

Implementations

impl CodecFrameFusedEngine

pub fn open( store: &Qwen3TtsWeightStore, talker_cfg: &TalkerConfig, cp_cfg: &CodePredictorConfig, session_device: Device, ) -> Result<CodecFrameFusedEngine, Error>

pub fn open_at( model_dir: &Path, store: &Qwen3TtsWeightStore, talker_cfg: &TalkerConfig, cp_cfg: &CodePredictorConfig, session_device: Device, ) -> Result<CodecFrameFusedEngine, Error>

pub fn device(&self) -> Device

pub fn warmup( &mut self, talker: &TalkerEngine, horizon: usize, ) -> Result<(), Error>

pub fn predict_codec_groups( &mut self, talker_hidden: &[f32], group0: u32, pad: &[f32], codec_emb: &mut [f32], ) -> Result<Vec<u32>, Error>

Host CP greedy (groups + summed codec embed); no talker decode.

pub fn run_codec_frame( &mut self, talker: &mut TalkerEngine, talker_hidden: &[f32], group0: u32, pad: &[f32], codec_emb: &mut [f32], hidden_out: &mut [f32], ) -> Result<Vec<u32>, Error>

Production path: host CP greedy once, then compiled talker decode.

pub fn run_talk_decode( &mut self, talker: &mut TalkerEngine, codec_emb: &[f32], hidden_out: &mut [f32], ) -> Result<(), Error>

Talker decode via TalkerEngine bucketed cache (no duplicate graph build per frame).

pub fn run_full_megagraph( &mut self, talker: &mut TalkerEngine, talker_hidden: &[f32], codec_emb: &[f32], step_embeds: &[Vec<f32>], hidden_out: &mut [f32], ) -> Result<(), Error>

Full megagraph (CP + talker); parity/bench only — duplicates CP vs Self::run_codec_frame.

pub fn set_cp_prefill_g0_embed(&mut self, g0_embed: &[f32])

Fill CP prefill row-1 with group-0 codec embed (megagraph feeds).

pub fn codec_embed_row( &self, group_idx: usize, token: u32, ) -> Result<Vec<f32>, Error>

pub fn cp_step_embeds_from_groups( &self, groups: &[u32], ) -> Result<Vec<Vec<f32>>, Error>