pub struct PrimeSlabs {Show 16 fields
pub t_cap: usize,
pub h: CudaSlice<f32>,
pub x1: CudaSlice<f32>,
pub z: CudaSlice<f32>,
pub act: CudaSlice<f32>,
pub xa: CudaSlice<f32>,
pub xb: CudaSlice<f32>,
pub h16: CudaSlice<u8>,
pub z16: CudaSlice<u8>,
pub gate: CudaSlice<f32>,
pub up: CudaSlice<f32>,
pub ffn_out: CudaSlice<f32>,
pub seg_glue: Vec<Option<CudaGraph>>,
pub mixed: CudaSlice<f32>,
pub seg_mid: Vec<Option<CudaGraph>>,
pub seg_t: usize,
}Expand description
Resident trunk transients for the eager prime (piecewise-graph foundation; see HybridModel::prime_slabs). Every buffer is fully overwritten before use per prime.
Fields§
§t_cap: usize§h: CudaSlice<f32>§x1: CudaSlice<f32>§z: CudaSlice<f32>§act: CudaSlice<f32>§xa: CudaSlice<f32>§xb: CudaSlice<f32>§h16: CudaSlice<u8>§z16: CudaSlice<u8>§gate: CudaSlice<f32>piecewise boundary slabs (increment 2): GEMM outputs land here so the downstream captured segments see fixed addresses.
up: CudaSlice<f32>§ffn_out: CudaSlice<f32>§seg_glue: Vec<Option<CudaGraph>>piecewise increment 3: per-layer S-glue segment graphs (down-add + next attn-norm, ALL-slab IO, zero in-graph allocations -> keeperless capture is clean). Baked at this t_cap; replay only when t == t_cap. seg_glue[il] fires between layer il and il+1 (ping-pong parity is deterministic per il).
mixed: CudaSlice<f32>increment 5 (core-split edition): the mixer out-GEMM writes into mixed
directly (no staging copy — the increment-4 copy route was refuted), making
S-mid [add + post-norm] all-slab and capturable.
seg_mid: Vec<Option<CudaGraph>>§seg_t: usize