Struct Gemma4E4bModel 

pub struct Gemma4E4bModel {
    pub tok_tbl_gpu: OnceLock<CudaSlice<u8>>,
    pub tok_embd_bytes: Vec<u8>,
    pub tok_embd_qt: i32,
    pub tok_embd_row_bytes: usize,
    pub model_proj: GpuTensor,
    pub proj_norm: GpuTensor,
    pub n_epl: usize,
}

gemma-4 E4B model-level per-layer-embedding tensors (prologue inputs). The token table stays HOST-side raw GGUF bytes at load (Q6_K [n_epl*n_layer, n_vocab], ~2.3GB VRAM when uploaded — the forward arc decides resident-vs-gather placement).

Fields

tok_tbl_gpu: OnceLock<CudaSlice<u8>>

device copy of the per-layer token table, uploaded on first use (the 26B embd_gpu pattern — keeps the ~2.3GB off load-critical paths that never decode).

tok_embd_bytes: Vec<u8>

tok_embd_qt: i32

tok_embd_row_bytes: usize

model_proj: GpuTensor

proj_norm: GpuTensor

n_epl: usize