Struct GemmaDraft 

pub struct GemmaDraft {

    pub layers: Vec<GemmaDraftLayer>,
    pub pre_proj: GpuTensor,
    pub post_proj: GpuTensor,
    pub output_norm: GpuTensor,
    pub head: GpuTensor,
    pub d2t: Option<Vec<u32>>,
    pub d2t_dev: Option<CudaSlice<u32>>,
    pub trim_adapt: Option<TrimAdapt>,
    pub rope_freqs: CudaSlice<f32>,
    pub ones: CudaSlice<f32>,
    pub n_embd: usize,
    pub n_backbone: usize,
    pub rope_base_global: f32,
    pub rope_base_swa: f32,
    pub sliding_window: usize,
}

Fields

layers: Vec<GemmaDraftLayer>

pre_proj: GpuTensor

post_proj: GpuTensor

output_norm: GpuTensor

head: GpuTensor

d2t: Option<Vec<u32>>

FR-Spec trim map: draft-row index -> target token id (None = full head, identity).

d2t_dev: Option<CudaSlice<u32>>

Device copy of d2t — the async round translates each drafted trim-idx in place (u32_map_k) before it seeds the next draft step or meets the verify argmax.

trim_adapt: Option<TrimAdapt>

Adaptive trim (coverage escapes are the entire trim cost — oracle-proven +2% on the cell the static trim lost by 17%, jsonl 2026-07-19): spare head slots learned at serve time from the prompt’s own ids and verify-correction tokens.

rope_freqs: CudaSlice<f32>

ones: CudaSlice<f32>

n_embd: usize

n_backbone: usize

rope_base_global: f32

rope_base_swa: f32

sliding_window: usize

Implementations

impl GemmaDraft

pub fn trim_adapt_stats(&self) -> Option<(usize, usize)>

Adaptive-trim stats: (slots used, slot budget). None when adaptation is off.

pub fn trim_adapt_save(&self) -> Result<usize>

Persist the learned trim rows: append ids not yet in the sidecar to <ranks>.learned (the load path pre-fills spare slots from it, so a distribution’s escapes pay their first-miss round ONCE across the serve lifetime, not per request).

impl GemmaDraft

pub fn load(e: &Engine, g: &GgufFile) -> Result<Self, Box<dyn Error>>