Struct EngramLayerWeights 

pub struct EngramLayerWeights {
    pub layer: usize,
    pub wkv: DevicePtr,
    pub qk: DevicePtr,
    pub raw: DevicePtr,
    pub rows: DevicePtr,
    pub wkv_q2k: DevicePtr,
}

2026-09-25: One engram layer’s resident weights on the device.

Fields

layer: usize

wkv: DevicePtr

2026-09-25: [dim * (hc + 1), cols * head_dim] bf16.

qk: DevicePtr

2026-09-25: [hc, dim] f32, the elementwise product q_weight * k_weight.

raw: DevicePtr

2026-09-25: This layer’s row workspace (add_layer allocates it): the raw Q2_K rows and their bf16 expansion, so every engram layer’s rows for a step can be resident at once.

rows: DevicePtr

wkv_q2k: DevicePtr

2026-09-25: wkv as the GGUF ships it, [out, in / 256] Q2_K blocks of ENGRAM_ROW_BYTES (84 bytes per 256 weights, against 512 in bf16). The single-token projection reads it instead of wkv; the two GEMVs agree bit for bit (checked on a GPU by gpu_engram_wkv_q2k_gemv_is_bitwise_the_bf16_gemv_over_the_expansion). Null: bf16 only.