Struct QLinear 

pub struct QLinear<'a> {
    pub out_dim: usize,
    pub in_dim: usize,
    pub weight: &'a [u32],
    pub scales: &'a [bf16],
    pub biases: &'a [bf16],
}

Zero-copy view of one quantized linear layer’s tensors.

Fields

out_dim: usize

in_dim: usize

weight: &'a [u32]

scales: &'a [bf16]

biases: &'a [bf16]

Implementations

impl<'a> QLinear<'a>

pub fn dequant_row(&self, row: usize, dst: &mut [f32])

Dequantize one output row into dst (len in_dim). Used for embedding lookup and for testing.

pub fn matvec(&self, x: &[f32], y: &mut [f32])

y = W x (f32 accumulate), parallel over output rows. Grouped form: y[r] = sum_g scale[r,g] * dot(q[r,g], x_g) + bias[r,g] * sum(x_g).