Function quantized_matmul 

pub fn quantized_matmul(
    x: &Array,
    w: &Array,
    scales: &Array,
    biases: &Array,
    transpose: bool,
    group_size: usize,
    bits: u32,
) -> Result<Array>

The matrix product of x with quantized weights, x · dequantize(w)ᵀ if transpose (MLX’s default, for weights stored as [out, in]) or x · dequantize(w), without materialising the dequantized matrix.

As in MLX, x, scales and biases are promoted to one floating-point type, and batch dimensions broadcast when x and w both have more than two. A one-dimensional x is a row vector.