Enum Linear 

pub enum Linear<B: Backend> {
    Dense(Tensor<B, 2>),
    Quant(Box<dyn QuantLinearOp<B>>),
}

A linear layer weight: dense tensor (portable path) or packed quant blocks bound to a fused device kernel.

Variants

Dense(Tensor<B, 2>)

Portable path: [n_out, k] dense tensor, burn matmul.

Quant(Box<dyn QuantLinearOp<B>>)

Fast path: packed weight + backend-specific kernel dispatch.

Implementations

impl<B: Backend> Linear<B>

pub fn dims(&self) -> [usize; 2]

[n_out, k].

pub fn forward( &self, x: Tensor<B, 3>, bias: Option<&Tensor<B, 1>>, ) -> Tensor<B, 3>

y = x @ W^T (+ b) for x: [batch, seq, k].