Struct EngramV41 

pub struct EngramV41 {
    pub dim: usize,
    pub hc: usize,
    pub head_dim: usize,
    pub cols: usize,
    pub eps: f32,
    pub max_tokens: usize,
    /* private fields */
}

The GPU side: kernels, per-layer weights, and workspaces for up to max_tokens positions per call.

Fields

dim: usize

hc: usize

head_dim: usize

cols: usize

eps: f32

max_tokens: usize

Implementations

impl EngramV41

pub fn apply( &self, gpu: &dyn GpuBackend, layer: usize, streams: DevicePtr, tokens: usize, stream: u64, ) -> Result<()>

Project the rows already in the workspace and gate tokens positions of streams ([T, hc, dim] f32) in place, for model layer layer.

impl EngramV41

pub fn new( gpu: &dyn GpuBackend, dim: usize, hc: usize, head_dim: usize, cols: usize, eps: f32, max_tokens: usize, ) -> Result<Self>

pub fn in_features(&self) -> usize

pub fn out_features(&self) -> usize

pub fn add_layer( &mut self, gpu: &dyn GpuBackend, w: EngramLayerWeights, ) -> Result<()>

Register a layer’s weights; its row workspace is allocated here.

pub fn layer(&self, layer: usize) -> Option<&EngramLayerWeights>

pub fn upload_qk( gpu: &dyn GpuBackend, q: &[f32], k: &[f32], ) -> Result<DevicePtr>

q_weight * k_weight (both [hc, dim], given as f32 values of the stored bf16) uploaded as the gate’s qk.

pub fn rows_from_q2k( &self, gpu: &dyn GpuBackend, layer: Option<usize>, raw_blocks: &[u8], n_rows: usize, stream: u64, ) -> Result<()>

Fill the row input from raw Q2_K blocks (n_rows * 84 bytes, one block per row in [token][col] order): upload, dequantise on the device.

pub fn rows_from_bf16( &self, gpu: &dyn GpuBackend, layer: Option<usize>, rows_bf16: &[u16], n_rows: usize, ) -> Result<()>

Fill the row input from already-dequantised bf16 rows (bits), for oracles and for tables held resident in bf16.

pub fn rows_ptr(&self) -> DevicePtr

The dequantised rows as the GEMM sees them ([n_rows, head_dim] bf16): the shared workspace (layer = None).

pub fn kv_ptr(&self) -> DevicePtr

The projected [tokens, dim * (hc + 1)] bf16 buffer after apply.

pub fn free(self, gpu: &dyn GpuBackend) -> Result<()>