Struct Linear 

pub struct Linear { /* private fields */ }

A row-major y = x · Wᵀ projection.

Quantized projections stay in the existing mmap-backed Proj/QTensor path, including Q8_2f and Q4TP. Native F32/F16/BF16 projections retain an Arc to the CMF mapping and are decoded into a single bounded scratch row.

Implementations

impl Linear

pub fn load(model: &Arc<CmfModel>, name: &str) -> Result<Self, String>

Load a rank-2 projection without materializing a whole F16/BF16 matrix. The CMF shape is semantic [out_features, in_features].

pub fn rows(&self) -> usize

pub fn cols(&self) -> usize

pub fn dtype(&self) -> Option<TensorDtype>

pub fn forward( &self, x: &[f32], batch: usize, out: &mut [f32], pool: Option<&Pool>, ) -> Result<(), String>

Compute out[b, rows] = x[b, cols] · Wᵀ.

The output and activation buffers are caller-owned. Mapped dense weights are decoded in bounded row tiles and sent through the existing GEMM path, so a 3,072×3,072 BF16 projection never becomes a 36 MiB temporary. Quantized projections use the existing pooled CPU/GPU dispatch and preserve Q8_2f/Q4TP semantics.

pub fn forward_one( &self, x: &[f32], out: &mut [f32], pool: Option<&Pool>, ) -> Result<(), String>

Compute one row, useful for small control projections and fixtures.

Trait Implementations

impl Debug for Linear

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.