Struct FcdModel 

pub struct FcdModel {
    pub hidden: usize,
    pub nh: usize,
    pub nkv: usize,
    pub hd: usize,
    pub nl: usize,
    pub vocab: usize,
    pub polish_only: bool,
    /* private fields */
}

The f32 training replica of a .cmf model (≤ 1B targets).

Fields

hidden: usize

nh: usize

nkv: usize

hd: usize

nl: usize

vocab: usize

polish_only: bool

Train the flagged layers WITHOUT switching their attention to the Nyström kernel — see FcdHyper::polish_only.

Implementations

impl FcdModel

pub fn from_cmf( model: &Arc<CmfModel>, o1: &O1Cfg, polish_only: bool, ) -> Result<Self, String>

Dequantize a model into the f32 training replica. Refuses what the backward cannot honestly differentiate yet (loud, not silent).

pub fn converted(&self) -> Vec<usize>

Converted (trainable) layer indices.

impl FcdModel

pub fn val_ppl( &self, va: &[u32], ts: Option<&TrainState>, student: bool, bs: usize, nrounds: usize, seq: usize, ) -> f64

Teacher-forced CE perplexity on deterministic evenly-spaced val windows (heal_hybridk_06b.py::val_ppl discipline — random windows made gate comparisons ride ±15% noise).

pub fn val_ppl_stream( &self, va: &FcdStream<'_>, ts: Option<&TrainState>, student: bool, bs: usize, nrounds: usize, seq: usize, ) -> f64

Self::val_ppl over an FcdStream: a plain stream takes the original evenly-spaced windows; a record stream scores up to 4·nrounds·bs records, evenly spaced, each window at its record’s start (or centred on its first scored token when the prompt is longer than half a window), weighted by the stream’s loss weights.