Skip to main content

blockscaled_gemm_int8

logicaffeine_compile::codegen_cuda::matrix

Function blockscaled_gemm_int8 

Source 
pub fn blockscaled_gemm_int8(
    m: usize,
    k: usize,
    n: usize,
    wq: &[i32],
    ws: &[f32],
    aq: &[i32],
    as_: &[f32],
) -> Vec<f64>
Expand description

WHAT THE DP4A KERNEL COMPUTES: an exact int32 dot within each 32-block, scaled once by the two block scales, accumulated in f64. Algebraically identical to blockscaled_gemm_dequant (the scale is constant across a block) — the lock proves the implementation matches, catching any scale-indexing bug.