Struct EngramLayerWeights 

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

One engram layer’s resident weights on the device.

Fields

layer: usize

wkv: DevicePtr

[dim * (hc + 1), cols * head_dim] bf16.

qk: DevicePtr

[hc, dim] f32, the elementwise product q_weight * k_weight.

raw: DevicePtr

This layer’s row workspace (add_layer allocates): the raw Q2_K rows and their bf16 expansion, so every engram layer’s rows for a step can be resident at once (the whole-step graph uploads them all up front).

rows: DevicePtr

wkv_q2k: DevicePtr

wkv as the GGUF ships it, [out, in / 256] raw Q2_K blocks (84 B), read by the single-token projection in place of the bf16 expansion (6x fewer bytes a token, byte-identical results); null = bf16 only.